fix(agentic): capture GB300 SGLang worker metrics

Layer topology-aware worker endpoint discovery onto the pinned srt-slurm release and opt all GB300 DeepSeek-V4 AgentX recipes into AIPerf server metrics collection. Add the missing GB300 hardware metadata so the DRAM-offload matrix generates.

中文：在固定版本的 srt-slurm 上叠加拓扑感知的 worker 指标端点发现逻辑，并为全部 GB300 DeepSeek-V4 AgentX 配方启用 AIPerf 服务端指标采集。同时补充 GB300 硬件元数据，确保 DRAM KV 卸载矩阵可正常生成。

Author: cquil11

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml

@@ -127,6 +127,7 @@ srun_options:
 
 benchmark:
   type: custom
+  aiperf_server_metrics: true
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1024-mtp-kvoffload.yaml

@@ -200,6 +200,7 @@ srun_options:
 
 benchmark:
   type: custom
+  aiperf_server_metrics: true
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c8-mtp-kvoffload.yaml

@@ -199,6 +199,7 @@ srun_options:
 
 benchmark:
   type: custom
+  aiperf_server_metrics: true
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml

@@ -200,6 +200,7 @@ srun_options:
 
 benchmark:
   type: custom
+  aiperf_server_metrics: true
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c256-mtp-kvoffload.yaml

@@ -200,6 +200,7 @@ srun_options:
 
 benchmark:
   type: custom
+  aiperf_server_metrics: true
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c512-mtp-kvoffload.yaml

@@ -200,6 +200,7 @@ srun_options:
 
 benchmark:
   type: custom
+  aiperf_server_metrics: true
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace

configs/runners.yaml

@@ -315,6 +315,9 @@ hardware:
   cluster:gb200-nv:
     available-cpu-dram-mib: 860_160
     gpus-per-node: 4
+  cluster:gb300-nv:
+    available-cpu-dram-mib: 860_160
+    gpus-per-node: 4
   cluster:mi300x-amds:
     available-cpu-dram-mib: 2_321_924
     gpus-per-node: 8

perf-changelog.yaml

@@ -4974,3 +4974,11 @@
     - "Run 29651235293 showed the 1M-context corpus working set outgrowing the HBM KV pool past conc 8 (TP8) / conc 64 (DP8): gpu_kv_cache_usage pinned at 1.0 and the radix hit rate collapsed from a ~0.97 theoretical ceiling to 0.04-0.06, so every post-knee turn re-prefilled its full history and throughput fell together with interactivity"
     - "HiCache spills evicted prefixes to host DRAM and restores them at C2C bandwidth instead of recomputing; sizing follows the qwen3.5-fp8-b300-sglang-agentic-hicache recipe (GLM-5.2 is plain GQA: one host pool per rank, GB-based --hicache-size)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2280
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-agg
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
+  description:
+    - "Correctly capture AIPerf Prometheus metrics from the Dynamo frontend and every logical SGLang worker endpoint"
+    - "Register GB300 hardware metadata required to generate AgentX DRAM-offload configs"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

runners/launch_gb300-nv.sh

@@ -142,10 +142,16 @@ if [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX ==
     # DSv4 GB300 sglang agentic: NVIDIA/srt-slurm v1.0.10 has the nginx
     # client_max_body_size fix (>1 MiB agentic warmup bodies), the
     # session-affinity frontend, and the BenchmarkType.CUSTOM / extra_mount
-    # schema these recipes need.
+    # schema these recipes need. Layer the pinned custom-AIPerf metrics patch
+    # onto that release so srt-slurm injects one topology-ordered Prometheus
+    # URL per logical SGLang worker leader. This keeps hostname/port discovery
+    # in the host-side orchestrator, which owns the resolved process topology.
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
     git checkout v1.0.10
+    git fetch https://github.com/cquil11/srt-slurm-nv.git \
+        de59739b172e507e15ebf145bfe305f606e82fbf
+    git cherry-pick --no-commit FETCH_HEAD
     mkdir -p recipes/sglang/deepseek-v4/agentic
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic" \
         recipes/sglang/deepseek-v4/agentic