update

Author: ch-wan

.github/configs/nvidia-master.yaml

@@ -8186,86 +8186,86 @@ dsv4-fp4-gb300-dynamo-sglang:
       osl: 1024
       search-space:
       # WideEP TP=16 decode: 1p1d-dep4-dep16. 5 nodes (4P + 16D = 20 GPUs).
-      - conc-list: [512]
+      - conc-list: [8192]
         prefill:
           num-worker: 1
           tp: 4
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc512-20.yaml"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-dep16-5-c8192.yaml"
         decode:
           num-worker: 1
           tp: 16
           ep: 16
           dp-attn: true
-      # DP-attn wideep: 1p1d-dep4-dep8. 3 nodes.
-      - conc-list: [512]
+      # WideEP TP=16 decode: 4p1d-dep4-dep16. 8 nodes.
+      - conc-list: [1024]
         prefill:
-          num-worker: 1
+          num-worker: 4
           tp: 4
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc512.yaml"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb300-4p1d-dep4-dep16-8-c1024.yaml"
         decode:
           num-worker: 1
-          tp: 8
-          ep: 8
+          tp: 16
+          ep: 16
           dp-attn: true
-      # DP-attn wideep: 2p1d-dep4-dep8. 4 nodes.
-      - conc-list: [1024]
+      # WideEP TP=16 decode: 8p1d-dep4-dep16. 12 nodes.
+      - conc-list: [4096]
         prefill:
-          num-worker: 2
+          num-worker: 8
           tp: 4
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc1024.yaml"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb300-8p1d-dep4-dep16-12-c4096.yaml"
         decode:
           num-worker: 1
-          tp: 8
-          ep: 8
+          tp: 16
+          ep: 16
           dp-attn: true
-      # Low concurrency
+      # Low concurrency: 1p1d-tp4-tp4. 2 nodes.
       - conc-list: [1]
         prefill:
           num-worker: 1
           tp: 4
           ep: 1
           dp-attn: false
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc1.yaml"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-tp4-tp4-2-c1.yaml"
         decode:
           num-worker: 1
           tp: 4
           ep: 1
           dp-attn: false
-      # Mid concurrency
-      - conc-list: [2048]
+      # Mid concurrency: 10p1d-dep4-dep16. 14 nodes.
+      - conc-list: [8192]
         prefill:
-          num-worker: 4
+          num-worker: 10
           tp: 4
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc2048.yaml"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb300-10p1d-dep4-dep16-14-c8192.yaml"
         decode:
           num-worker: 1
-          tp: 8
-          ep: 8
+          tp: 16
+          ep: 16
           dp-attn: true
-      # Max concurrency
-      - conc-list: [16384]
+      # Max concurrency: 12p1d-dep4-dep12. 15 nodes.
+      - conc-list: [21504]
         prefill:
-          num-worker: 14
+          num-worker: 12
           tp: 4
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc16384.yaml"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb300-12p1d-dep4-dep12-15-c21504.yaml"
         decode:
           num-worker: 1
-          tp: 16
-          ep: 16
+          tp: 12
+          ep: 12
           dp-attn: true

…es/sglang/deepseek-v4/8k1k/conc2048.yaml …agg-gb300-10p1d-dep4-dep16-14-c8192.yamlbenchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc2048.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-10p1d-dep4-dep16-14-c8192.yaml benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc2048.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-10p1d-dep4-dep16-14-c8192.yaml

@@ -1,4 +1,4 @@
-name: "conc2048"
+name: "disagg-gb300-10p1d-dep4-dep16-14-c8192"
 
 # 8k/1k high-throughput topology for the wideep DSV4-Pro setup.
 #
@@ -37,7 +37,7 @@ model:
   precision: "fp4"
 
 dynamo:
-  hash: "9d3c913d300eb368cda28b3f98a23a5762621e0d"
+  hash: "81d0555ee23519cea80a42b4fe824e30368b7300"
   install: true
 
 slurm:
@@ -186,6 +186,7 @@ backend:
       max-running-requests: 21504
       cuda-graph-max-bs: 1280
 
+
 benchmark:
   type: "sa-bench"
   isl: 8192

…s/sglang/deepseek-v4/8k1k/conc16384.yaml …gg-gb300-12p1d-dep4-dep12-15-c21504.yamlbenchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc16384.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-12p1d-dep4-dep12-15-c21504.yaml benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc16384.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-12p1d-dep4-dep12-15-c21504.yaml

@@ -1,4 +1,4 @@
-name: "conc16384"
+name: "disagg-gb300-12p1d-dep4-dep12-15-c21504"
 
 # 8k/1k high-throughput topology for the wideep DSV4-Pro setup.
 #
@@ -37,7 +37,7 @@ model:
   precision: "fp4"
 
 dynamo:
-  hash: "9d3c913d300eb368cda28b3f98a23a5762621e0d"
+  hash: "81d0555ee23519cea80a42b4fe824e30368b7300"
   install: true
 
 slurm:
@@ -175,7 +175,7 @@ backend:
       disaggregation-transfer-backend: mooncake
       disaggregation-decode-polling-interval: 8
 
-      mem-fraction-static: 0.962
+      mem-fraction-static: 0.94
       swa-full-tokens-ratio: 0.056
       context-length: 9216
       tensor-parallel-size: 12

…/sglang/deepseek-v4/8k1k/conc512-20.yaml …isagg-gb300-1p1d-dep4-dep16-5-c8192.yamlbenchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc512-20.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-dep16-5-c8192.yaml benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc512-20.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-dep16-5-c8192.yaml

@@ -1,4 +1,4 @@
-name: "conc16384"
+name: "disagg-gb300-1p1d-dep4-dep16-5-c8192"
 
 # 8k/1k high-throughput topology for the wideep DSV4-Pro setup.
 #
@@ -37,7 +37,7 @@ model:
   precision: "fp4"
 
 dynamo:
-  hash: "9d3c913d300eb368cda28b3f98a23a5762621e0d"
+  hash: "81d0555ee23519cea80a42b4fe824e30368b7300"
   install: true
 
 slurm:
@@ -175,7 +175,7 @@ backend:
       disaggregation-transfer-backend: mooncake
       disaggregation-decode-polling-interval: 8
 
-      mem-fraction-static: 0.962
+      mem-fraction-static: 0.94
       swa-full-tokens-ratio: 0.056
       context-length: 9216
       tensor-parallel-size: 16

…cipes/sglang/deepseek-v4/8k1k/conc1.yaml …8k1k/disagg-gb300-1p1d-tp4-tp4-2-c1.yamlbenchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc1.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-tp4-tp4-2-c1.yaml benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc1.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-tp4-tp4-2-c1.yaml

@@ -1,4 +1,4 @@
-name: "conc1"
+name: "disagg-gb300-1p1d-tp4-tp4-2-c1"
 
 # 8k/1k high-throughput topology for the wideep DSV4-Pro setup.
 #
@@ -40,7 +40,7 @@ model:
 # rationale. Hash bumped from PR #1213 to track the dynamo-sglang dsv4
 # dev branch.
 dynamo:
-  hash: "9d3c913d300eb368cda28b3f98a23a5762621e0d"
+  hash: "81d0555ee23519cea80a42b4fe824e30368b7300"
   install: true
 
 slurm:

…pes/sglang/deepseek-v4/8k1k/conc512.yaml …isagg-gb300-4p1d-dep4-dep16-8-c1024.yamlbenchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc512.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-4p1d-dep4-dep16-8-c1024.yaml benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc512.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-4p1d-dep4-dep16-8-c1024.yaml

@@ -1,4 +1,4 @@
-name: "conc16384"
+name: "disagg-gb300-4p1d-dep4-dep16-8-c1024"
 
 # 8k/1k high-throughput topology for the wideep DSV4-Pro setup.
 #
@@ -37,7 +37,7 @@ model:
   precision: "fp4"
 
 dynamo:
-  hash: "9d3c913d300eb368cda28b3f98a23a5762621e0d"
+  hash: "81d0555ee23519cea80a42b4fe824e30368b7300"
   install: true
 
 slurm:
@@ -175,7 +175,7 @@ backend:
       disaggregation-transfer-backend: mooncake
       disaggregation-decode-polling-interval: 8
 
-      mem-fraction-static: 0.962
+      mem-fraction-static: 0.94
       swa-full-tokens-ratio: 0.056
       context-length: 9216
       tensor-parallel-size: 16

…es/sglang/deepseek-v4/8k1k/conc1024.yaml …sagg-gb300-8p1d-dep4-dep16-12-c4096.yamlbenchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc1024.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-8p1d-dep4-dep16-12-c4096.yaml benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc1024.yaml renamed to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-8p1d-dep4-dep16-12-c4096.yaml

@@ -1,4 +1,4 @@
-name: "conc16384"
+name: "disagg-gb300-8p1d-dep4-dep16-12-c4096"
 
 # 8k/1k high-throughput topology for the wideep DSV4-Pro setup.
 #
@@ -37,7 +37,7 @@ model:
   precision: "fp4"
 
 dynamo:
-  hash: "9d3c913d300eb368cda28b3f98a23a5762621e0d"
+  hash: "81d0555ee23519cea80a42b4fe824e30368b7300"
   install: true
 
 slurm:
@@ -175,7 +175,7 @@ backend:
       disaggregation-transfer-backend: mooncake
       disaggregation-decode-polling-interval: 8
 
-      mem-fraction-static: 0.962
+      mem-fraction-static: 0.94
       swa-full-tokens-ratio: 0.056
       context-length: 9216
       tensor-parallel-size: 16

utils/matrix_logic/generate_sweep_configs.py

@@ -115,11 +115,7 @@ def _max_eval_conc(ie):
         )
         mn_groups[key].append((i, entry))
 
-    for key, entries in mn_groups.items():
-        # TODO(pr1157): srt-slurm pin (9d75f82) lacks the lm-eval orchestrator path
-        # (only on sa-submission-q2-2026). Skip eval-only here until the pin is bumped.
-        if key[:3] == ("deepseek-ai/DeepSeek-V4-Pro", "gb300-cw", "dynamo-sglang"):
-            continue
+    for entries in mn_groups.values():
         best_idx, best_entry = max(entries, key=_max_eval_conc)
         eval_indices.add(best_idx)
         # Set eval-conc to median of eligible conc values to avoid OOM during eval