DeepSeek-V4.1-Flash H200 FP4 SGLang AgentX: move to nightly-20261013-7224fce1 and TP4 c2 max-running-requests 2 / DeepSeek-V4.1-Flash H200 FP4 SGLang AgentX：升级到 nightly-20261013-7224fce1 并将 TP4 c2 max-running-requests 设为 2

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/h200-fp4-mtp/agentic.yaml

@@ -67,9 +67,7 @@ base:
       AIPERF_HTTP_TCP_USER_TIMEOUT: '900000'
       AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
 
-# One variant per point. Admission is 2x CONC, capped at the decode graph batch;
-# saturation points carry a larger in-flight working set than the 30-minute
-# warmup drain allows.
+# One variant per point. Admission is 2x CONC, capped at the decode graph batch.
 override_tp4_c1:
   roles:
     agg:
@@ -88,7 +86,7 @@ override_tp4_c2:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        max-running-requests: 4
+        max-running-requests: 2
   benchmark:
     env:
       CONC: '2'
@@ -130,42 +128,6 @@ override_tp4_c16:
       CONC: '16'
       AGENTIC_WARMUP_GRACE_PERIOD: '1800'
 
-override_tp4_c32:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        tensor-parallel-size: 4
-        max-running-requests: 64
-  benchmark:
-    env:
-      CONC: '32'
-      AGENTIC_WARMUP_GRACE_PERIOD: '3600'
-
-override_tp4_c64:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        tensor-parallel-size: 4
-        max-running-requests: 64
-  benchmark:
-    env:
-      CONC: '64'
-      AGENTIC_WARMUP_GRACE_PERIOD: '3600'
-
-override_tp4_c128:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        tensor-parallel-size: 4
-        max-running-requests: 64
-  benchmark:
-    env:
-      CONC: '128'
-      AGENTIC_WARMUP_GRACE_PERIOD: '3600'
-
 override_tp8_c1:
   roles:
     agg:
@@ -225,39 +187,3 @@ override_tp8_c16:
     env:
       CONC: '16'
       AGENTIC_WARMUP_GRACE_PERIOD: '1800'
-
-override_tp8_c32:
-  roles:
-    agg:
-      gpus: 8
-      args:
-        tensor-parallel-size: 8
-        max-running-requests: 64
-  benchmark:
-    env:
-      CONC: '32'
-      AGENTIC_WARMUP_GRACE_PERIOD: '3600'
-
-override_tp8_c64:
-  roles:
-    agg:
-      gpus: 8
-      args:
-        tensor-parallel-size: 8
-        max-running-requests: 64
-  benchmark:
-    env:
-      CONC: '64'
-      AGENTIC_WARMUP_GRACE_PERIOD: '3600'
-
-override_tp8_c128:
-  roles:
-    agg:
-      gpus: 8
-      args:
-        tensor-parallel-size: 8
-        max-running-requests: 64
-  benchmark:
-    env:
-      CONC: '128'
-      AGENTIC_WARMUP_GRACE_PERIOD: '3600'

inferencex-e2e/configs/nvidia-master.yaml

@@ -8617,7 +8617,7 @@ dsv41flash-fp4-h200-vllm-agentic-dspark:
 # Match published vLLM TP4/EP1 and TP8/EP1 using dsv4 attention and the nightly's
 # native MXFP4 x BF16 Marlin MoE runner.
 dsv41flash-fp4-h200-sglang-agentic-dspark:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260923-06008c17
+  image: lmsysorg/sglang:nightly-dev-cu13-20261013-7224fce1
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:h200-dgxc
@@ -8629,8 +8629,8 @@ dsv41flash-fp4-h200-sglang-agentic-dspark:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 4, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: agentic.yaml }
-      - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: agentic.yaml }
+      - { tp: 4, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16], srt-recipe: agentic.yaml }
+      - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16], srt-recipe: agentic.yaml }
 
 dsv4-fp4-b200-dynamo-sglang-agentic-agg:
   image: lmsysorg/sglang:nightly-dev-20260916-c9a8fba9

inferencex-e2e/perf-changelog.yaml

@@ -10259,6 +10259,15 @@
     - "No change to the Inferact/Kimi-K3-DSpark draft's precision: the checkpoint is all BF16 with no quantization_config, and at both images the draft takes its own quantization config (get_draft_quant_config()), so the target quantization-config does not reach it. The draft runs BF16 weights with an fp8 KV cache."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4215
 
+- config-keys:
+    - dsv41flash-fp4-h200-sglang-agentic-dspark
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Update the image from lmsysorg/sglang:nightly-dev-cu13-20260923-06008c17 to lmsysorg/sglang:nightly-dev-cu13-20261013-7224fce1."
+    - "Set TP4 concurrency 2 max-running-requests to 2 and sweep concurrency 1 to 16 on TP4 and TP8."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
 - config-keys:
     - dsv4-fp4-mi355x-sglang-agentic-mtp
   scenario-type: