perf(dsv41flash): add B200 vLLM AgentX DEP2 c24, DEP4 c96 and DEP2 c64 host-KV rows / 新增 B200 vLLM AgentX DEP2 c24、DEP4 c96 与 DEP2 c64 主机 KV 行

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

Author: hshrivastava-droid

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

@@ -127,7 +127,7 @@ zip_override_dep2:
 
 # DEP4: TP1 x DP4 + EP4 with MegaMoE behind a consistent-hash router. At
 # c128, new prefills are admitted every 48 engine steps, aligned across DP
-# ranks.
+# ranks; c64 keeps the vLLM default of 1.
 zip_override_dep4:
   frontend:
     type: vllm-router
@@ -147,12 +147,12 @@ zip_override_dep4:
         enable-expert-parallel: true
         kernel-config: '{"enable_flashinfer_autotune":true,"moe_backend":"deep_gemm_mega_moe"}'
         engram-config: '{"cpu_offload":true,"embedding_across_dp":true}'
-        max-num-seqs: [128]
-        prefill-schedule-interval: [48]
+        max-num-seqs: [64, 128]
+        prefill-schedule-interval: [1, 48]
         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],"decoder_replay_cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2048,2304,2560,2816,3072,3328,3584,3840,4096]}'
   benchmark:
     env:
-      CONC: ['128']
+      CONC: ['64', '128']
 
 # TP2 c1: the TP4 c1 settings on two GPUs.
 override_tp2_c1:
@@ -205,3 +205,93 @@ override_dep2_c128_dram:
     env:
       CONC: '128'
       KV_OFFLOADING: dram
+
+# DEP2 c24: the DEP2 c32 settings (piecewise graphs to 4096 tokens,
+# prefill-schedule-interval 64) at concurrency 24.
+override_dep2_c24:
+  frontend:
+    type: vllm-router
+    args:
+      policy: consistent_hash
+      prometheus-host: 127.0.0.1
+      prometheus-port: 18000
+      request-timeout-secs: 14400
+      disable-retries: true
+  setup_script: vllm-router-0.1.14.sh
+  roles:
+    agg:
+      gpus: 2
+      args:
+        tensor-parallel-size: 1
+        data-parallel-size: 2
+        enable-expert-parallel: true
+        kernel-config: '{"enable_flashinfer_autotune":true,"moe_backend":"deep_gemm_mega_moe"}'
+        engram-config: '{"cpu_offload":true,"embedding_across_dp":true}'
+        max-num-seqs: 32
+        compilation-config: '{"mode":"VLLM_COMPILE","cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,640,768,896,1024,1280,1536,1792,2048,2304,2560,2816,3072,3328,3584,3840,4096]}'
+        max-cudagraph-capture-size: 4096
+        max-num-batched-tokens: 4096
+        prefill-schedule-interval: 64
+  benchmark:
+    env:
+      CONC: '24'
+
+# DEP4 c96: the DEP4 c128 settings (prefill-schedule-interval 48) at
+# concurrency 96.
+override_dep4_c96:
+  frontend:
+    type: vllm-router
+    args:
+      policy: consistent_hash
+      prometheus-host: 127.0.0.1
+      prometheus-port: 18000
+      request-timeout-secs: 14400
+      disable-retries: true
+  setup_script: vllm-router-0.1.14.sh
+  roles:
+    agg:
+      gpus: 4
+      args:
+        tensor-parallel-size: 1
+        data-parallel-size: 4
+        enable-expert-parallel: true
+        kernel-config: '{"enable_flashinfer_autotune":true,"moe_backend":"deep_gemm_mega_moe"}'
+        engram-config: '{"cpu_offload":true,"embedding_across_dp":true}'
+        max-num-seqs: 96
+        prefill-schedule-interval: 48
+        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],"decoder_replay_cudagraph_capture_sizes":[6,12,18,24,30,36,48,60,72,96,120,144,192,240,288,384,480,576,768,1020,1536,2048,2304,2560,2816,3072,3328,3584,3840,4096]}'
+  benchmark:
+    env:
+      CONC: '96'
+
+# DEP2 c64 host KV: the DEP2 c128 host-KV settings at concurrency 64.
+override_dep2_c64_dram:
+  frontend:
+    type: vllm-router
+    args:
+      policy: consistent_hash
+      prometheus-host: 127.0.0.1
+      prometheus-port: 18000
+      request-timeout-secs: 14400
+      disable-retries: true
+  setup_script: vllm-router-0.1.14.sh
+  roles:
+    agg:
+      gpus: 2
+      args:
+        tensor-parallel-size: 1
+        data-parallel-size: 2
+        enable-expert-parallel: true
+        kernel-config: '{"enable_flashinfer_autotune":true,"moe_backend":"deep_gemm_mega_moe"}'
+        engram-config: '{"cpu_offload":true,"embedding_across_dp":true}'
+        max-num-seqs: 48
+        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]}'
+        max-cudagraph-capture-size: 576
+        max-num-batched-tokens: 4096
+        gpu-memory-utilization: 0.955
+        kv-offloading-backend: native
+        kv-offloading-size: 32
+  benchmark:
+    env:
+      CONC: '64'
+      KV_OFFLOADING: dram

inferencex-e2e/configs/nvidia-master.yaml

@@ -8673,11 +8673,11 @@ dsv41flash-fp4-b200-vllm-agentic-dspark:
       # TP2 runs the TP4 c1 settings on two GPUs.
       - { tp: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b200-fp4-mtp/agentic.yaml }
       # DEP2 is native TP1 x DP2 + EP2 with MegaMoE behind a consistent-hash vLLM Router.
-      - { tp: 2, ep: 2, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [8, 16, 32], router: { name: vllm-router, version: "0.1.14" }, srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b200-fp4-mtp/agentic.yaml }
+      - { tp: 2, ep: 2, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [8, 16, 24, 32], router: { name: vllm-router, version: "0.1.14" }, srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b200-fp4-mtp/agentic.yaml }
       # DEP2 c128 adds vLLM's native CPU KV offloading at 32 GiB per DP rank.
-      - { tp: 2, ep: 2, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: native }, spec-decoding: mtp, conc-list: [128], router: { name: vllm-router, version: "0.1.14" }, srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b200-fp4-mtp/agentic.yaml }
+      - { tp: 2, ep: 2, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: native }, spec-decoding: mtp, conc-list: [64, 128], router: { name: vllm-router, version: "0.1.14" }, srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b200-fp4-mtp/agentic.yaml }
       # DEP4 is native TP1 x DP4 + EP4 with MegaMoE behind the same router.
-      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [128], router: { name: vllm-router, version: "0.1.14" }, srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b200-fp4-mtp/agentic.yaml }
+      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [64, 96, 128], router: { name: vllm-router, version: "0.1.14" }, srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b200-fp4-mtp/agentic.yaml }
 
 # SGLang arm for DeepSeek-V4.1-Flash AgentX on B200, from the SGLang cookbook
 # (https://lmsysorg.mintlify.app/cookbook/autoregressive/DeepSeek/DeepSeek-V4_1).

inferencex-e2e/perf-changelog.yaml

@@ -9652,6 +9652,12 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Tune B200 vLLM AgentX DEP2/DEP4 with prefill-schedule-interval and extended piecewise CUDA graph capture."
-    - "Add TP2 c1 and DEP2 c128 host-KV (native CPU offloading) rows."
+    - "DEP4 c128: admit new prefills every 48 engine steps (prefill-schedule-interval 48). DEP4 c64 is unchanged."
+    - "DEP2 c32: extend piecewise CUDA graph capture to 4096 tokens (max-cudagraph-capture-size 4096) and set prefill-schedule-interval 64. DEP2 c8 and c16 are unchanged."
+    - "Add a TP2 c1 row that runs the TP4 c1 settings on two GPUs."
+    - "Add a DEP2 c24 row with the DEP2 c32 settings (piecewise graphs to 4096 tokens, prefill-schedule-interval 64) at concurrency 24."
+    - "Add a DEP4 c96 row with the DEP4 c128 settings (prefill-schedule-interval 48) and max-num-seqs 96."
+    - "Add a DEP2 c64 host-KV row with the DEP2 c128 host-KV settings at concurrency 64."
+    - "Add a DEP2 c128 row with vLLM native CPU KV offloading (kv-offloading: dram, backend native), 32 GiB per DP rank, gpu-memory-utilization 0.955 and max-num-seqs 48."
+    - "Image, DSpark settings and all other points are unchanged."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3837