[AgentX] Update Muse Spark Open GB300 SGLang image and retune concurrency points / [AgentX] 更新 Muse Spark Open GB300 SGLang 镜像并调整并发点

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/musespark/sglang/gb300-fp4-mtp/agentic.yaml

@@ -43,13 +43,15 @@ base:
         tool-call-parser: muse
         watchdog-timeout: 3600
         enable-metrics: true
+        stream-interval: 20
       env:
         PYTHONNOUSERSITE: '1'
         PYTHONUNBUFFERED: '1'
         SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: '1'
         # Outlast AIPerf's pooled connections past Uvicorn's keep-alive.
         SGLANG_TIMEOUT_KEEP_ALIVE: '900'
         HF_HUB_OFFLINE: '1'
+        SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION: '0'
   # Cold 533 GB weight loads from shared storage plus graph capture outlast the default.
   health_check:
     max_attempts: 1440
@@ -103,8 +105,8 @@ override_tp4_c10:
   roles:
     agg:
       args:
-        max-running-requests: 10
-        cuda-graph-max-bs-decode: 10
+        max-running-requests: 3
+        cuda-graph-max-bs-decode: 3
         mem-fraction-static: 0.88
         prefill-decode-interval: 64
         speculative-draft-attention-backend: fa4
@@ -129,39 +131,49 @@ override_tp4_c15:
     env:
       CONC: '15'
 
-override_tp4_c16:
+override_tp4_c24:
   roles:
     agg:
       args:
-        max-running-requests: 12
-        cuda-graph-max-bs-decode: 12
+        max-running-requests: 48
+        cuda-graph-max-bs-decode: 48
         mem-fraction-static: 0.95
+        swa-full-tokens-ratio: 0.1
+        cuda-graph-backend-prefill: breakable
+        max-prefill-tokens: 8192
         schedule-policy: shortest-prefill-first
-        prefill-decode-interval: 48
+        prefill-decode-interval: 96
         enable-hierarchical-cache: true
         hicache-size: 120
         hicache-write-policy: write_through
+      env:
+        SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '2048'
   benchmark:
     env:
-      CONC: '16'
+      CONC: '24'
 
-override_tp4_c20:
+override_tp4_c28:
   roles:
     agg:
       args:
-        max-running-requests: 10
-        cuda-graph-max-bs-decode: 10
+        max-running-requests: 48
+        cuda-graph-max-bs-decode: 48
         mem-fraction-static: 0.95
+        swa-full-tokens-ratio: 0.1
+        cuda-graph-backend-prefill: breakable
+        max-prefill-tokens: 8192
         schedule-policy: shortest-prefill-first
-        prefill-decode-interval: 48
+        prefill-decode-interval: 32
         enable-hierarchical-cache: true
-        hicache-size: 120
-        hicache-write-policy: write_through
+        hicache-size: 160
+        hicache-write-policy: write_back
+      env:
+        SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '2048'
   benchmark:
     env:
-      CONC: '20'
+      CONC: '28'
 
-override_tp4_c24:
+override_tp4_c30:
   roles:
     agg:
       args:
@@ -172,15 +184,15 @@ override_tp4_c24:
         cuda-graph-backend-prefill: breakable
         max-prefill-tokens: 8192
         schedule-policy: shortest-prefill-first
-        prefill-decode-interval: 32
+        prefill-decode-interval: 64
         enable-hierarchical-cache: true
-        hicache-size: 120
-        hicache-write-policy: write_through
+        hicache-size: 160
+        hicache-write-policy: write_back
       env:
         SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '2048'
   benchmark:
     env:
-      CONC: '24'
+      CONC: '30'
 
 override_tp4_c34:
   roles:
@@ -196,16 +208,16 @@ override_tp4_c34:
         schedule-policy: shortest-prefill-first
         prefill-decode-interval: 48
         enable-hierarchical-cache: true
-        hicache-size: 150
-        hicache-write-policy: write_through
+        hicache-size: 160
+        hicache-write-policy: write_back
       env:
         SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '4096'
         SGLANG_TRTLLM_MHA_DECODE_SEQ_LEN_SPLITS: '4'
   benchmark:
     env:
       CONC: '34'
 
-override_tp4_c40:
+override_tp4_c44:
   roles:
     agg:
       args:
@@ -218,11 +230,33 @@ override_tp4_c40:
         speculative-draft-kv-cache-dtype: fp8_e4m3
         enable-hierarchical-cache: true
         hicache-size: 149
-        hicache-write-policy: write_through
+        hicache-write-policy: write_back
+        tokenizer-worker-num: 4
+        incremental-streaming-output: true
+      env:
+        SGLANG_TRTLLM_MHA_DECODE_SEQ_LEN_SPLITS: '4'
+  benchmark:
+    env:
+      CONC: '44'
+
+override_tp4_c48:
+  roles:
+    agg:
+      args:
+        max-running-requests: 24
+        cuda-graph-max-bs-decode: 24
+        mem-fraction-static: 0.9
+        swa-full-tokens-ratio: 0.1
+        chunked-prefill-size: null
+        speculative-draft-attention-backend: trtllm_mha
+        speculative-draft-kv-cache-dtype: fp8_e4m3
+        enable-hierarchical-cache: true
+        hicache-size: 149
+        hicache-write-policy: write_back
         tokenizer-worker-num: 4
         incremental-streaming-output: true
       env:
         SGLANG_TRTLLM_MHA_DECODE_SEQ_LEN_SPLITS: '4'
   benchmark:
     env:
-      CONC: '40'
+      CONC: '48'

inferencex-e2e/configs/nvidia-master.yaml

@@ -9558,7 +9558,7 @@ qwen3.8next-fp4-gb300-sglang-agentic-mtp:
         srt-recipe: agentic.yaml
 
 musespark-fp4-gb300-sglang-agentic-dflash:
-  image: lmsysorg/sglang:dev-muse-spark-open@sha256:8cfc5a72cb296771f5f8a4a4688cb6d05685610a830cbb616fa4ab0eb7bb2b0a
+  image: lmsysorg/sglang:nightly-dev-20261021-c1c287a4
   model: meta-models/Muse-Spark-Open-816B-A42B-NVFP4
   model-prefix: musespark
   runner: cluster:gb300-nv
@@ -9572,7 +9572,7 @@ musespark-fp4-gb300-sglang-agentic-dflash:
       - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: draft_model, conc-list: [1, 5, 8, 10, 15], srt-recipe: agentic.yaml }
     - dram-utilization: 0.80
       search-space:
-      - { tp: 4, ep: 4, kv-offloading: dram, kv-offload-backend: { name: hicache }, spec-decoding: draft_model, conc-list: [16, 20, 24, 34, 40], srt-recipe: agentic.yaml }
+      - { tp: 4, ep: 4, kv-offloading: dram, kv-offload-backend: { name: hicache }, spec-decoding: draft_model, conc-list: [24, 28, 30, 34, 44, 48], srt-recipe: agentic.yaml }
 
 musespark-fp4-gb300-dynamo-sglang-agentic-dflash:
   image: lmsysorg/sglang:dev-muse-spark-open@sha256:8cfc5a72cb296771f5f8a4a4688cb6d05685610a830cbb616fa4ab0eb7bb2b0a

inferencex-e2e/perf-changelog.yaml

@@ -10779,6 +10779,15 @@
     - "Enable SGLANG_HIP_FP4_INDEXER_ROWGROUP=1 on all points. Enable SGLANG_HIP_DSV41_INDEX_K_PAGE8=1 only on the non-monokernel points (TP2 c32, DP2 c32 and c64); the monokernel refuses that layout."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4367
 
+- config-keys:
+    - musespark-fp4-gb300-sglang-agentic-dflash
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Update the Muse Spark Open GB300 curve to lmsysorg/sglang:nightly-dev-20261021-c1c287a4 and set SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 and --stream-interval 20 at every concurrency."
+    - "Lower max-running-requests to 3 at concurrency 10, raise prefill-decode-interval to 96 at concurrency 24, switch HiCache to write_back, and replace concurrency 16, 20 and 40 with 28, 30, 44 and 48."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
 - config-keys:
     - musespark-fp4-b200-sglang-agentic-dflash
   scenario-type: