chokevin · chokevin · Apr 20, 2026 · Apr 20, 2026
diff --git a/docs/profiling/INDEX.md 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 |
+|---|---|---|---|---|---|---|---|
+| 20260420T015907Z | `d212cd6` | voice | fp16,marlin | NVIDIA A100-SXM4-80GB | 0.7 | [`20260420T015907Z/`](./marlin/20260420T015907Z/) | [link](https://github.com/chokevin/swordfish/pull/3) |
diff --git a/docs/profiling/marlin/20260420T015907Z/70b-tp2-b1.ncu-csv.log b/docs/profiling/marlin/20260420T015907Z/70b-tp2-b1.ncu-csv.log
diff --git a/docs/profiling/marlin/20260420T015907Z/70b-tp2-b1.ncu.csv b/docs/profiling/marlin/20260420T015907Z/70b-tp2-b1.ncu.csv
@@ -0,0 +1,10 @@
+==PROF== Connected to process 907 (/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 907
+==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/20260420T015907Z/70b-tp2-b1.ncu.log b/docs/profiling/marlin/20260420T015907Z/70b-tp2-b1.ncu.log
@@ -0,0 +1,2 @@
+==ERROR== Option '--section SpeedOfLight --section SpeedOfLight_RooflineChart --section MemoryWorkloadAnalysis --section SchedulerStats --section WarpStateStats --section ComputeWorkloadAnalysis' did not match any section.
+==ERROR== Use --list-sections to see the list of available sections.
diff --git a/docs/profiling/marlin/20260420T015907Z/70b-tp2-b4.ncu-csv.log b/docs/profiling/marlin/20260420T015907Z/70b-tp2-b4.ncu-csv.log
diff --git a/docs/profiling/marlin/20260420T015907Z/70b-tp2-b4.ncu.csv b/docs/profiling/marlin/20260420T015907Z/70b-tp2-b4.ncu.csv
@@ -0,0 +1,10 @@
+==PROF== Connected to process 996 (/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 996
+==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/20260420T015907Z/70b-tp2-b4.ncu.log b/docs/profiling/marlin/20260420T015907Z/70b-tp2-b4.ncu.log
@@ -0,0 +1,2 @@
+==ERROR== Option '--section SpeedOfLight --section SpeedOfLight_RooflineChart --section MemoryWorkloadAnalysis --section SchedulerStats --section WarpStateStats --section ComputeWorkloadAnalysis' did not match any section.
+==ERROR== Use --list-sections to see the list of available sections.
diff --git a/docs/profiling/marlin/20260420T015907Z/70b-tp2-b8.ncu-csv.log b/docs/profiling/marlin/20260420T015907Z/70b-tp2-b8.ncu-csv.log
diff --git a/docs/profiling/marlin/20260420T015907Z/70b-tp2-b8.ncu.csv b/docs/profiling/marlin/20260420T015907Z/70b-tp2-b8.ncu.csv
@@ -0,0 +1,10 @@
+==PROF== Connected to process 1085 (/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 1085
+==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/20260420T015907Z/70b-tp2-b8.ncu.log b/docs/profiling/marlin/20260420T015907Z/70b-tp2-b8.ncu.log
@@ -0,0 +1,2 @@
+==ERROR== Option '--section SpeedOfLight --section SpeedOfLight_RooflineChart --section MemoryWorkloadAnalysis --section SchedulerStats --section WarpStateStats --section ComputeWorkloadAnalysis' did not match any section.
+==ERROR== Use --list-sections to see the list of available sections.
diff --git a/docs/profiling/marlin/20260420T015907Z/8b-b1.ncu-csv.log b/docs/profiling/marlin/20260420T015907Z/8b-b1.ncu-csv.log
diff --git a/docs/profiling/marlin/20260420T015907Z/8b-b1.ncu.csv b/docs/profiling/marlin/20260420T015907Z/8b-b1.ncu.csv
@@ -0,0 +1,10 @@
+==PROF== Connected to process 640 (/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 640
+==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/20260420T015907Z/8b-b1.ncu.log b/docs/profiling/marlin/20260420T015907Z/8b-b1.ncu.log
@@ -0,0 +1,2 @@
+==ERROR== Option '--section SpeedOfLight --section SpeedOfLight_RooflineChart --section MemoryWorkloadAnalysis --section SchedulerStats --section WarpStateStats --section ComputeWorkloadAnalysis' did not match any section.
+==ERROR== Use --list-sections to see the list of available sections.
diff --git a/docs/profiling/marlin/20260420T015907Z/8b-b4.ncu-csv.log b/docs/profiling/marlin/20260420T015907Z/8b-b4.ncu-csv.log
diff --git a/docs/profiling/marlin/20260420T015907Z/8b-b4.ncu.csv b/docs/profiling/marlin/20260420T015907Z/8b-b4.ncu.csv
@@ -0,0 +1,10 @@
+==PROF== Connected to process 729 (/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 729
+==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/20260420T015907Z/8b-b4.ncu.log b/docs/profiling/marlin/20260420T015907Z/8b-b4.ncu.log
@@ -0,0 +1,2 @@
+==ERROR== Option '--section SpeedOfLight --section SpeedOfLight_RooflineChart --section MemoryWorkloadAnalysis --section SchedulerStats --section WarpStateStats --section ComputeWorkloadAnalysis' did not match any section.
+==ERROR== Use --list-sections to see the list of available sections.
diff --git a/docs/profiling/marlin/20260420T015907Z/8b-b8.ncu-csv.log b/docs/profiling/marlin/20260420T015907Z/8b-b8.ncu-csv.log
diff --git a/docs/profiling/marlin/20260420T015907Z/8b-b8.ncu.csv b/docs/profiling/marlin/20260420T015907Z/8b-b8.ncu.csv
@@ -0,0 +1,10 @@
+==PROF== Connected to process 818 (/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 818
+==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/20260420T015907Z/8b-b8.ncu.log b/docs/profiling/marlin/20260420T015907Z/8b-b8.ncu.log
@@ -0,0 +1,2 @@
+==ERROR== Option '--section SpeedOfLight --section SpeedOfLight_RooflineChart --section MemoryWorkloadAnalysis --section SchedulerStats --section WarpStateStats --section ComputeWorkloadAnalysis' did not match any section.
+==ERROR== Use --list-sections to see the list of available sections.
diff --git a/docs/profiling/marlin/20260420T015907Z/SUMMARY.md b/docs/profiling/marlin/20260420T015907Z/SUMMARY.md
@@ -0,0 +1,26 @@
+# Autoresearch run `20260420T015907Z`
+
+- **source SHA:** `d212cd6`
+- **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 | 0.032 | 0.034 | 1.1 | x1.00 | |
+| 8b-b1 | marlin | 0.049 | 0.050 | 0.7 | x0.65 | |
+| 8b-b4 | fp16 | 0.031 | 0.032 | 4.3 | x1.00 | |
+| 8b-b4 | marlin | 0.049 | 0.050 | 2.7 | x0.64 | |
+| 8b-b8 | fp16 | 0.032 | 0.032 | 8.5 | x1.00 | |
+| 8b-b8 | marlin | 0.049 | 0.050 | 5.4 | x0.64 | |
+| 70b-tp2-b1 | fp16 | 0.051 | 0.059 | 1.3 | x1.00 | |
+| 70b-tp2-b1 | marlin | 0.050 | 0.051 | 1.4 | x1.03 | |
+| 70b-tp2-b4 | fp16 | 0.049 | 0.050 | 5.4 | x1.00 | |
+| 70b-tp2-b4 | marlin | 0.049 | 0.049 | 5.5 | x1.01 | |
+| 70b-tp2-b8 | fp16 | 0.050 | 0.050 | 10.8 | x1.00 | |
+| 70b-tp2-b8 | marlin | 0.048 | 0.049 | 11.1 | x1.02 | |
+
+![roofline](./roofline.png)
diff --git a/docs/profiling/marlin/20260420T015907Z/env.txt b/docs/profiling/marlin/20260420T015907Z/env.txt
@@ -0,0 +1,43 @@
+=== profile_marlin.sh @ 20260420T015907Z ===
+--- host ---
+Linux swordfish-profile-sf-prof-260420-015138-7xlss 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 01:59:07 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             69W /  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 ---
+d212cd669477b14cb33a352af64ae27dee669634
diff --git a/docs/profiling/marlin/20260420T015907Z/manifest.json b/docs/profiling/marlin/20260420T015907Z/manifest.json
@@ -0,0 +1,223 @@
+{
+  "env": {
+    "host": "swordfish-profile-sf-prof-260420-015138-7xlss",
+    "os": "Linux 6.6.126.1-1.azl3",
+    "python": "3.10.12",
+    "torch": "2.4.0a0+07cecf4168.nv24.05",
+    "cuda_available": true,
+    "timestamp": "2026-04-20T01:59:13+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": 0.03189593601226807,
+      "ms_p50": 0.03144511938095093,
+      "ms_p95": 0.033960320949554444,
+      "ms_min": 0.030894720554351808,
+      "tflops_mean": 1.051997094146854,
+      "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": 0.049192959785461426,
+      "ms_p50": 0.04912576198577881,
+      "ms_p95": 0.04963967800140381,
+      "ms_min": 0.04897664070129395,
+      "tflops_mean": 0.6820982544318616,
+      "error": null,
+      "speedup_vs_fp16": 0.6483841621112346
+    },
+    {
+      "name": "8b-b4",
+      "M": 4,
+      "N": 4096,
+      "K": 4096,
+      "group_size": 128,
+      "priority": 0,
+      "tag": "llama-3-8b",
+      "impl": "fp16",
+      "ms_mean": 0.03144793605804443,
+      "ms_p50": 0.03157952070236206,
+      "ms_p95": 0.03182080030441284,
+      "ms_min": 0.031113600730895995,
+      "tflops_mean": 4.267934396466279,
+      "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": 0.04927884769439697,
+      "ms_p50": 0.04931519985198975,
+      "ms_p95": 0.04962495803833008,
+      "ms_min": 0.048726401329040527,
+      "tflops_mean": 2.7236377123172995,
+      "error": null,
+      "speedup_vs_fp16": 0.6381629751789037
+    },
+    {
+      "name": "8b-b8",
+      "M": 8,
+      "N": 4096,
+      "K": 4096,
+      "group_size": 128,
+      "priority": 0,
+      "tag": "llama-3-8b",
+      "impl": "fp16",
+      "ms_mean": 0.03154764842987061,
+      "ms_p50": 0.031646080017089843,
+      "ms_p95": 0.03203840017318726,
+      "ms_min": 0.03096640110015869,
+      "tflops_mean": 8.508889548351702,
+      "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": 0.04929036808013916,
+      "ms_p50": 0.0494700813293457,
+      "ms_p95": 0.04981311798095703,
+      "ms_min": 0.04864255905151367,
+      "tflops_mean": 5.44600226079793,
+      "error": null,
+      "speedup_vs_fp16": 0.6400367791650207
+    },
+    {
+      "name": "70b-tp2-b1",
+      "M": 1,
+      "N": 8192,
+      "K": 4096,
+      "group_size": 128,
+      "priority": 0,
+      "tag": "llama-3-70b",
+      "impl": "fp16",
+      "ms_mean": 0.051108479499816895,
+      "ms_p50": 0.049178881645202635,
+      "ms_p95": 0.059057278633117674,
+      "ms_min": 0.04889984130859375,
+      "tflops_mean": 1.3130671203051625,
+      "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": 0.04959603214263916,
+      "ms_p50": 0.049487361907958986,
+      "ms_p95": 0.05057280063629151,
+      "ms_min": 0.04831935882568359,
+      "tflops_mean": 1.353109535194138,
+      "error": null,
+      "speedup_vs_fp16": 1.0304953298043662
+    },
+    {
+      "name": "70b-tp2-b4",
+      "M": 4,
+      "N": 8192,
+      "K": 4096,
+      "group_size": 128,
+      "priority": 0,
+      "tag": "llama-3-70b",
+      "impl": "fp16",
+      "ms_mean": 0.04933452701568604,
+      "ms_p50": 0.04925183773040771,
+      "ms_p95": 0.04956480026245117,
+      "ms_min": 0.04922239780426026,
+      "tflops_mean": 5.4411275882841705,
+      "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": 0.04871065616607666,
+      "ms_p50": 0.04873280048370361,
+      "ms_p95": 0.04931136131286621,
+      "ms_min": 0.04817728042602539,
+      "tflops_mean": 5.510815848687854,
+      "error": null,
+      "speedup_vs_fp16": 1.0128076872436766
+    },
+    {
+      "name": "70b-tp2-b8",
+      "M": 8,
+      "N": 8192,
+      "K": 4096,
+      "group_size": 128,
+      "priority": 0,
+      "tag": "llama-3-70b",
+      "impl": "fp16",
+      "ms_mean": 0.04958604812622071,
+      "ms_p50": 0.04956863880157471,
+      "ms_p95": 0.049699201583862304,
+      "ms_min": 0.04951680183410644,
+      "tflops_mean": 10.827055841058382,
+      "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": 0.0484769287109375,
+      "ms_p50": 0.048438401222229005,
+      "ms_p95": 0.048990721702575686,
+      "ms_min": 0.04785344123840332,
+      "tflops_mean": 11.074771572293722,
+      "error": null,
+      "speedup_vs_fp16": 1.0228793251712947
+    }
+  ]
+}
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1,2 @@
		==ERROR== Option '--section SpeedOfLight --section SpeedOfLight_RooflineChart --section MemoryWorkloadAnalysis --section SchedulerStats --section WarpStateStats --section ComputeWorkloadAnalysis' did not match any section.
		==ERROR== Use --list-sections to see the list of available sections.