Update GB300 DeepSeek V4.1 Flash vLLM nightly / 更新 GB300 DeepSeek V4.1 Flash vLLM nightly

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/gb300-fp4-mtp/agentic.yaml

@@ -6,7 +6,7 @@ base:
   name: dsv41flash-fp4-gb300-vllm-agentic
   model:
     path: hf:deepseek-ai/DeepSeek-V4.1-Flash
-    container: vllm/vllm-openai:nightly-af1c01499b289be555c475669ba50a88e96d846e
+    container: vllm/vllm-openai:nightly-ac68c3087215e0a4f3cdfa218508c6aada57235d
     precision: fp4
   resources:
     gpu_type: gb300
@@ -37,12 +37,17 @@ base:
         enable-auto-tool-choice: true
         reasoning-parser: deepseek_v41
         engram-config: '{"cpu_offload":true}'
+        # FlashInfer sparse attention with MXFP4 indexer KV and sparse logits.
+        attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV41","indexer_kv_dtype":"mxfp4","indexer_sparse_logits":true}'
+        kv-cache-dtype: fp8
         # Five-token DSpark with probabilistic drafting. Throughput runs replace
         # block rejection with the golden acceptance length.
         speculative-config: '{"method":"dspark","num_speculative_tokens":5,"draft_sample_method":"probabilistic","rejection_sample_method":"block","enable_adaptive_verification":false}'
         max-model-len: 1048576
+        max-num-seqs: 256
         disable-uvicorn-access-log: true
       env:
+        VLLM_DISABLED_KERNELS: FlashInferCutedslMxfp8LinearKernel
         VLLM_USE_RUST_FRONTEND: '1'
         PYTHONUNBUFFERED: '1'
         VLLM_ENGINE_READY_TIMEOUT_S: '7200'
@@ -53,18 +58,20 @@ base:
       MODEL: deepseek-ai/DeepSeek-V4.1-Flash
       AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'vllm:'
 
-# One variant per point. Graph capture starts at 64 tokens and doubles until it
-# covers CONC x (1 + 5 drafts), up to 2048. TP2 leaves ~175 GiB of weights on
-# each 277 GiB GPU, 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.
+# One variant per point. Piecewise graph capture sizes are multiples of the
+# six-token DSpark verification block, denser for small batches, and each
+# batched-token limit matches the largest captured graph: CONC <= 4 and TP2
+# CONC 128 capture up to 2046 tokens (the latter at 0.97 memory utilization),
+# every other point up to 8190.
 override_tp4_c1:
   roles:
     agg:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        max-cudagraph-capture-size: 64
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046]}'
+        max-cudagraph-capture-size: 2046
+        max-num-batched-tokens: 2048
   benchmark:
     env:
       CONC: '1'
@@ -75,7 +82,9 @@ override_tp4_c2:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        max-cudagraph-capture-size: 64
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046]}'
+        max-cudagraph-capture-size: 2046
+        max-num-batched-tokens: 2048
   benchmark:
     env:
       CONC: '2'
@@ -86,7 +95,9 @@ override_tp4_c4:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        max-cudagraph-capture-size: 64
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046]}'
+        max-cudagraph-capture-size: 2046
+        max-num-batched-tokens: 2048
   benchmark:
     env:
       CONC: '4'
@@ -97,7 +108,9 @@ override_tp4_c8:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        max-cudagraph-capture-size: 64
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046,3072,4092,6144,8190]}'
+        max-cudagraph-capture-size: 8190
+        max-num-batched-tokens: 8192
   benchmark:
     env:
       CONC: '8'
@@ -108,7 +121,9 @@ override_tp4_c16:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        max-cudagraph-capture-size: 128
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046,3072,4092,6144,8190]}'
+        max-cudagraph-capture-size: 8190
+        max-num-batched-tokens: 8192
   benchmark:
     env:
       CONC: '16'
@@ -119,7 +134,9 @@ override_tp4_c32:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        max-cudagraph-capture-size: 256
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046,3072,4092,6144,8190]}'
+        max-cudagraph-capture-size: 8190
+        max-num-batched-tokens: 8192
   benchmark:
     env:
       CONC: '32'
@@ -130,7 +147,9 @@ override_tp4_c64:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        max-cudagraph-capture-size: 512
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046,3072,4092,6144,8190]}'
+        max-cudagraph-capture-size: 8190
+        max-num-batched-tokens: 8192
   benchmark:
     env:
       CONC: '64'
@@ -141,33 +160,22 @@ override_tp4_c128:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        max-cudagraph-capture-size: 1024
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046,3072,4092,6144,8190]}'
+        max-cudagraph-capture-size: 8190
+        max-num-batched-tokens: 8192
   benchmark:
     env:
       CONC: '128'
 
-override_tp2_c1:
-  roles:
-    agg:
-      gpus: 2
-      args:
-        tensor-parallel-size: 2
-        max-cudagraph-capture-size: 64
-        max-num-batched-tokens: 4096
-        max-num-seqs: 16
-  benchmark:
-    env:
-      CONC: '1'
-
 override_tp2_c2:
   roles:
     agg:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        max-cudagraph-capture-size: 64
-        max-num-batched-tokens: 4096
-        max-num-seqs: 16
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046]}'
+        max-cudagraph-capture-size: 2046
+        max-num-batched-tokens: 2048
   benchmark:
     env:
       CONC: '2'
@@ -178,9 +186,9 @@ override_tp2_c4:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        max-cudagraph-capture-size: 64
-        max-num-batched-tokens: 4096
-        max-num-seqs: 16
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046]}'
+        max-cudagraph-capture-size: 2046
+        max-num-batched-tokens: 2048
   benchmark:
     env:
       CONC: '4'
@@ -191,9 +199,9 @@ override_tp2_c8:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        max-cudagraph-capture-size: 64
-        max-num-batched-tokens: 4096
-        max-num-seqs: 16
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046,3072,4092,6144,8190]}'
+        max-cudagraph-capture-size: 8190
+        max-num-batched-tokens: 8192
   benchmark:
     env:
       CONC: '8'
@@ -204,9 +212,9 @@ override_tp2_c16:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        max-cudagraph-capture-size: 128
-        max-num-batched-tokens: 4096
-        max-num-seqs: 32
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046,3072,4092,6144,8190]}'
+        max-cudagraph-capture-size: 8190
+        max-num-batched-tokens: 8192
   benchmark:
     env:
       CONC: '16'
@@ -217,9 +225,9 @@ override_tp2_c32:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        max-cudagraph-capture-size: 256
-        max-num-batched-tokens: 4096
-        max-num-seqs: 64
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046,3072,4092,6144,8190]}'
+        max-cudagraph-capture-size: 8190
+        max-num-batched-tokens: 8192
   benchmark:
     env:
       CONC: '32'
@@ -230,9 +238,9 @@ override_tp2_c64:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        max-cudagraph-capture-size: 512
-        max-num-batched-tokens: 4096
-        max-num-seqs: 128
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046,3072,4092,6144,8190]}'
+        max-cudagraph-capture-size: 8190
+        max-num-batched-tokens: 8192
   benchmark:
     env:
       CONC: '64'
@@ -243,9 +251,23 @@ override_tp2_c128:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        max-cudagraph-capture-size: 512
-        max-num-batched-tokens: 4096
-        max-num-seqs: 256
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2046]}'
+        max-cudagraph-capture-size: 2046
+        max-num-batched-tokens: 2048
+        gpu-memory-utilization: 0.97
   benchmark:
     env:
       CONC: '128'
+
+override_tp2_c1:
+  roles:
+    agg:
+      gpus: 2
+      args:
+        tensor-parallel-size: 2
+        max-cudagraph-capture-size: 64
+        max-num-batched-tokens: 4096
+        max-num-seqs: 16
+  benchmark:
+    env:
+      CONC: '1'

inferencex-e2e/configs/nvidia-master.yaml

@@ -8384,7 +8384,7 @@ dsv41flash-fp4-b200-sglang-agentic-dspark:
       - { tp: 2, ep: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/b200-fp4-mtp/agentic.yaml }
 
 dsv41flash-fp4-gb300-vllm-agentic-dspark:
-  image: vllm/vllm-openai:nightly-af1c01499b289be555c475669ba50a88e96d846e
+  image: vllm/vllm-openai:nightly-ac68c3087215e0a4f3cdfa218508c6aada57235d
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:gb300-nv

inferencex-e2e/perf-changelog.yaml

@@ -9219,3 +9219,9 @@
   description:
     - "Update Kimi-K3 B200 AgentX to vllm/vllm-openai:nightly-ac9126e58aa7bbab1856ba6593ba4d5003fea516, set bfloat16 KDA state, and add concurrency 32/56/72/80 with TP8/PP2/DCP8 and DSpark K4."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3087
+
+- config-keys:
+    - dsv41flash-fp4-gb300-vllm-agentic-dspark
+  description:
+    - "Pin GB300 DeepSeek-V4.1-Flash vLLM to nightly ac68c308 with FlashInfer sparse attention, MXFP4 indexer KV, sparse logits, fp8 KV cache, and B300's batching and CUDA graph tiers; use FlashInfer CUTLASS for MXFP8 GEMMs to avoid the CuTe-DSL split-K startup failure."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3396