[Klaud Cold] Update dsv41flash-fp4-b200-vllm-agentic-dspark to vLLM nightly-36768d1b / [Klaud Cold] 将 dsv41flash-fp4-b200-vllm-agentic-dspark 更新至 vLLM nightly-36768d1b

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b200-fp4-mtp/agentic.yaml

@@ -6,7 +6,7 @@ base:
   name: dsv41flash-fp4-b200-vllm-agentic
   model:
     path: hf:deepseek-ai/DeepSeek-V4.1-Flash
-    container: vllm/vllm-openai:nightly-ddd6fbca148a867aad1fcab7ec72f582b9977db4
+    container: vllm/vllm-openai:nightly-36768d1bfd39094681cdbc8cb37d4b31c0729c89
     precision: fp4
   resources:
     gpu_type: b200
@@ -62,13 +62,17 @@ base:
 # rank, so it caps batched tokens at 4096 (the indexer's 1M-wide logits buffer),
 # bounds the scheduler at 2x CONC within [16, 256] (FlashInfer autotune fails
 # below 16) and stops capturing above 512 tokens.
+# TP4 enables the fused all-reduce + mHC MoE finalize, whose 16-token autotune
+# pass selects an invalid MXFP8 split-K tactic, so TP4 leaves mxfp8_gemm untuned.
 override_tp4_c1:
   roles:
     agg:
       gpus: 4
       args:
         tensor-parallel-size: 4
         max-cudagraph-capture-size: 64
+      env:
+        VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS: mxfp8_gemm
   benchmark:
     env:
       CONC: '1'
@@ -80,6 +84,8 @@ override_tp4_c2:
       args:
         tensor-parallel-size: 4
         max-cudagraph-capture-size: 64
+      env:
+        VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS: mxfp8_gemm
   benchmark:
     env:
       CONC: '2'
@@ -91,6 +97,8 @@ override_tp4_c4:
       args:
         tensor-parallel-size: 4
         max-cudagraph-capture-size: 64
+      env:
+        VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS: mxfp8_gemm
   benchmark:
     env:
       CONC: '4'
@@ -102,6 +110,8 @@ override_tp4_c8:
       args:
         tensor-parallel-size: 4
         max-cudagraph-capture-size: 64
+      env:
+        VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS: mxfp8_gemm
   benchmark:
     env:
       CONC: '8'
@@ -113,6 +123,8 @@ override_tp4_c16:
       args:
         tensor-parallel-size: 4
         max-cudagraph-capture-size: 128
+      env:
+        VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS: mxfp8_gemm
   benchmark:
     env:
       CONC: '16'
@@ -124,6 +136,8 @@ override_tp4_c32:
       args:
         tensor-parallel-size: 4
         max-cudagraph-capture-size: 256
+      env:
+        VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS: mxfp8_gemm
   benchmark:
     env:
       CONC: '32'
@@ -135,6 +149,8 @@ override_tp4_c64:
       args:
         tensor-parallel-size: 4
         max-cudagraph-capture-size: 512
+      env:
+        VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS: mxfp8_gemm
   benchmark:
     env:
       CONC: '64'
@@ -146,6 +162,8 @@ override_tp4_c128:
       args:
         tensor-parallel-size: 4
         max-cudagraph-capture-size: 1024
+      env:
+        VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS: mxfp8_gemm
   benchmark:
     env:
       CONC: '128'

inferencex-e2e/configs/nvidia-master.yaml

@@ -8358,7 +8358,7 @@ dsv41flash-fp4-b300-vllm-agentic-dspark:
       - { tp: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b300-fp4-mtp/agentic.yaml }
 
 dsv41flash-fp4-b200-vllm-agentic-dspark:
-  image: vllm/vllm-openai:nightly-ddd6fbca148a867aad1fcab7ec72f582b9977db4
+  image: vllm/vllm-openai:nightly-36768d1bfd39094681cdbc8cb37d4b31c0729c89
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:b200-nscale

inferencex-e2e/perf-changelog.yaml

@@ -9062,3 +9062,9 @@
     - "All points now run vllm/vllm-openai-rocm:v0.30.0 (previously v0.29.0)."
     - "LMCache DRAM point (TP8, concurrency 16): one LMCache server per node with 1296 GB L1 shared by all 8 TP ranks, replacing 8 per-rank servers of 162 GB L1 each."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3545
+
+- config-keys:
+    - dsv41flash-fp4-b200-vllm-agentic-dspark
+  description:
+    - "Update vLLM to nightly-36768d1b and skip TP4 MXFP8 GEMM autotuning to avoid an invalid split-K tactic."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3579