perf(agentx): refresh dsv4-fp4-b200-vllm-agentic-mtp harness

benchmarks/single_node/agentic/dsv4_fp4_b200_vllm_mtp.sh

@@ -93,6 +93,13 @@ if [ "$DP_ATTENTION" = "true" ]; then
     agentic_pip_install --quiet "vllm-router==$VLLM_ROUTER_VERSION"
 fi
 
+# AIPerf automatically scrapes the public endpoint's /metrics URL. That is the
+# vLLM engine for pure TP, but the native router for DP-attention. Explicitly
+# add the engine endpoint so every topology captures vLLM metrics; AIPerf
+# deduplicates it against the automatic endpoint in pure-TP runs.
+export AIPERF_SERVER_METRICS_URLS="http://localhost:${VLLM_BACKEND_PORT}/metrics"
+export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="vllm:"
+
 # DeepSeek-V4-Pro weights are large; engine startup can exceed default 600s.
 export VLLM_ENGINE_READY_TIMEOUT_S=3600
 

perf-changelog.yaml

@@ -5656,3 +5656,12 @@
     - "Search space now runs on the AgentX MTP concurrency grid: the GPU-resident row [4, 8, 12, 16] and the HiCache row [4, 8, 12, 16]. Every step is at least 2 concurrency apart and every arm stops hard at conc 16 -- single-step sampling cannot separate configurations by more than run-to-run noise on the agentic corpus, and past conc 16 these SKUs are into the post-HBM-cliff thrashing regime that the non-MTP sweeps already characterized, which is not worth one GPU job per point."
     - "Drops conc 1 from the GPU-resident arm. The TP8/EP8 c1 cell was the one job that never returned on this recipe while the other 18 completed, so it gated the PR without adding a point: at conc 1 a single session cannot exercise the EP8 all-to-all this arm exists to measure, and the low-interactivity corner is already covered by conc 4."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2428
+
+- config-keys:
+    - dsv4-fp4-b200-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Refresh submission with up to date AgentX harness."
+    - "Capture the vLLM backend Prometheus endpoint for every topology and require vllm: metrics before publishing AgentX results."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2516