diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep8-c128-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep8-c128-mtp-agentic.yaml index dd58dde28f..56879ed607 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep8-c128-mtp-agentic.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep8-c128-mtp-agentic.yaml @@ -195,6 +195,7 @@ benchmark: INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace" RESULT_DIR: "/logs/agentic" PORT: "8000" + AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:" IS_MULTINODE: "true" AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0" AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep8-dep8-c384-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep8-dep8-c384-mtp-agentic.yaml index 00a876ebe8..1c071050e7 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep8-dep8-c384-mtp-agentic.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep8-dep8-c384-mtp-agentic.yaml @@ -196,6 +196,7 @@ benchmark: INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace" RESULT_DIR: "/logs/agentic" PORT: "8000" + AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:" IS_MULTINODE: "true" AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0" AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true" diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 4a2145e445..4934640533 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5526,4 +5526,13 @@ - "Enable SGLang backend metrics on every aggregate, prefill, and decode engine, and fail before publishing a partial trace artifact if required sglang: metrics are absent." - "Record the recipes' active HiCache host-DRAM tier and, for disaggregated points, Dynamo router commit 5a638087 in nvidia-master metadata so generated artifacts no longer report kv_offloading=none, allocated_cpu_dram_gb=0, or a null router." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2477 + +- config-keys: + - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg + scenario-type: + - agentic-coding + description: + - "Refresh submission with up to date AgentX harness." + - "Require vLLM backend metrics in every AgentX artifact." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2519 diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh index d955d18ebc..10bd8df770 100644 --- a/runners/launch_gb300-nv.sh +++ b/runners/launch_gb300-nv.sh @@ -217,11 +217,13 @@ elif [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX = cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic" \ recipes/sglang/deepseek-v4/agentic elif [[ "$IS_AGENTIC" == "1" ]]; then - # Agentic recipes use NVIDIA/srt-slurm v1.0.36. This is the upstream - # version validated in InferenceX PR #2302 and includes per-node DP, - # matching Dynamo health counts, multi-node TP port handling, and - # Mooncake compatibility. Keep it pinned so sweeps are reproducible. - git clone --branch v1.0.36 --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR" || exit 1 + # Agentic recipes use NVIDIA/srt-slurm v1.0.38. In addition to per-node + # DP, matching Dynamo health counts, multi-node TP port handling, and + # Mooncake compatibility, this release injects logical worker endpoint + # metadata and AIPERF_SERVER_METRICS_URLS into custom benchmarks. The + # latter is required for complete AgentX server-metrics artifacts. + # Keep the release pinned so sweeps are reproducible. + git clone --branch v1.0.38 --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR" || exit 1 cd "$SRT_REPO_DIR" || exit 1 mkdir -p recipes/vllm/deepseek-v4/agentic || exit 1