[PowerX] Qwen3.5 B200 AgentX measured-power coverage

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/8k1k.yaml

@@ -56,9 +56,9 @@ zip_override_tp4_ep1:
     agg:
       gpus: 4
       args:
-        cuda-graph-max-bs-decode: [2, 4, 8, 16, 32, 64, 128, 256, 512, 320, 384, 448, 640]
-        scheduler-recv-interval: [10, 10, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30]
+        cuda-graph-max-bs-decode: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 320, 384, 448, 640]
+        scheduler-recv-interval: [10, 10, 10, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30]
         tensor-parallel-size: 4
   benchmark:
     env:
-      CONC: ['2', '4', '8', '16', '32', '64', '128', '256', '512', '320', '384', '448', '640']
+      CONC: ['1', '2', '4', '8', '16', '32', '64', '128', '256', '512', '320', '384', '448', '640']

inferencex-e2e/configs/nvidia-master.yaml

@@ -1142,7 +1142,7 @@ qwen3.5-fp8-b200-sglang:
         srt-recipe: 8k1k.yaml
       - tp: 4
         ep: 1
-        conc-start: 2
+        conc-start: 1
         conc-end: 512
         srt-recipe: 8k1k.yaml
       - tp: 4

inferencex-e2e/configs/powerx-article-b200.yaml

@@ -0,0 +1,19 @@
+qwen3.5-fp8-b200-sglang:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260918-20518d85
+  model: Qwen/Qwen3.5-397B-A17B-FP8
+  model-prefix: qwen3.5
+  runner: cluster:b200-nscale
+  precision: fp8
+  framework: sglang
+  multinode: false
+  srt-recipe-dir: qwen3.5/sglang/b200-fp8
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - tp: 4
+        ep: 1
+        conc-start: 1
+        conc-end: 128
+        srt-recipe: 8k1k.yaml

inferencex-e2e/perf-changelog.yaml

@@ -10052,3 +10052,11 @@
     - "Offload every Muse Spark Open H200 point to host DRAM with SGLang HiCache (120 GB per rank so the target and DFlash draft host pools fit the node cgroup, dram-utilization 0.80), keeping the concurrency 1-64 sweep."
     - "Keep the GPU-resident settings unchanged: mem-fraction-static 0.85, matching running-request and decode-graph caps."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4199
+
+- config-keys:
+    - qwen3.5-fp8-b200-sglang
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Add TP4/EP1 concurrency 1 to the Qwen3.5 FP8 B200 SGLang 8k1k recipe zip_override (cuda-graph-max-bs-decode 1, scheduler-recv-interval 10) so the c1 matrix point matches, and add the configs/powerx-article-b200.yaml selector for TP4/EP1 c1-c128 article repeats. The reused sweep measured and validated GPU power at every point."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3378