fix(gb300): correct vLLM worker startup checks

Backport per-node Dynamo health counting and provide explicit GB300 NUMA mappings for aggregate workers.

中文：回移按节点启动模式下的 Dynamo 健康检查计数修复，并为 GB300 聚合式工作进程显式配置 NUMA 映射。

Author: ivanium

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp4-agentic.yaml

@@ -104,6 +104,7 @@ backend:
       pipeline-parallel-size: 1
       enable-cumem-allocator: true
       numa-bind: true
+      numa-bind-nodes: [0, 0, 1, 1]
       attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
       max-model-len: 1048576
       max-num-seqs: 32

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp8-agentic.yaml

@@ -104,6 +104,7 @@ backend:
       pipeline-parallel-size: 1
       enable-cumem-allocator: true
       numa-bind: true
+      numa-bind-nodes: [0, 0, 1, 1]
       attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
       max-model-len: 1048576
       max-num-seqs: 32

benchmarks/multi_node/srt-slurm-vllm-per-node-health.patch

@@ -0,0 +1,63 @@
+diff --git a/src/srtctl/cli/mixins/benchmark_stage.py b/src/srtctl/cli/mixins/benchmark_stage.py
+index fe6cbf1..f282fe6 100644
+--- a/src/srtctl/cli/mixins/benchmark_stage.py
++++ b/src/srtctl/cli/mixins/benchmark_stage.py
+@@ -46,12 +46,32 @@ def _vllm_data_parallel_size(config: "SrtConfig", mode: str) -> int:
+     return int(mode_config.get("data-parallel-size") or mode_config.get("data_parallel_size") or 1)
+
+
+-def _get_health_expectations(config: "SrtConfig") -> tuple[int, int, str, int]:
++def _vllm_health_entries(
++    config: "SrtConfig",
++    mode: str,
++    logical_workers: int,
++    backend_processes: list["Process"] | None,
++) -> int:
++    """Return expected Dynamo generate registrations for a vLLM worker mode."""
++    dp_size = _vllm_data_parallel_size(config, mode)
++    if dp_size > 1 and getattr(config.backend, "dp_launch_mode", "per_gpu") == "per_node":
++        if backend_processes is None:
++            raise ValueError("backend_processes are required for per-node DP health expectations")
++        endpoint_mode = "agg" if mode == "aggregated" else mode
++        return sum(process.endpoint_mode == endpoint_mode for process in backend_processes)
++
++    return logical_workers * dp_size
++
++
++def _get_health_expectations(
++    config: "SrtConfig", backend_processes: list["Process"] | None = None
++) -> tuple[int, int, str, int]:
+     """Compute expected health counts in the units reported by the frontend.
+
+     Dynamo's /health endpoint reports registered generate instances. For vLLM
+-    DP workers, that means one entry per DP rank, not one entry per logical
+-    srt-slurm worker. Other frontends keep using logical worker counts.
++    DP workers, per-GPU launch registers one entry per DP rank, while per-node
++    launch registers one entry per node-local process. Other frontends keep
++    using logical worker counts.
+     """
+     r = config.resources
+
+@@ -67,10 +87,10 @@ def _get_health_expectations(config: "SrtConfig") -> tuple[int, int, str, int]:
+     if config.frontend.type == "dynamo" and getattr(config.backend, "type", None) == "vllm":
+         if r.num_agg > 0:
+             n_prefill = 0
+-            n_decode = logical_decode * _vllm_data_parallel_size(config, "aggregated")
++            n_decode = _vllm_health_entries(config, "aggregated", logical_decode, backend_processes)
+         else:
+-            n_prefill = logical_prefill * _vllm_data_parallel_size(config, "prefill")
+-            n_decode = logical_decode * _vllm_data_parallel_size(config, "decode")
++            n_prefill = _vllm_health_entries(config, "prefill", logical_prefill, backend_processes)
++            n_decode = _vllm_health_entries(config, "decode", logical_decode, backend_processes)
+
+         count_desc = f"{n_prefill}P + {n_decode}D Dynamo generate instances; logical workers: {worker_desc}"
+         return n_prefill, n_decode, count_desc, n_prefill + n_decode
+@@ -131,7 +151,7 @@ class BenchmarkStageMixin:
+         """Run the benchmark."""
+         logger.info("Waiting for workers to be ready...")
+
+-        n_prefill, n_decode, count_desc, num_workers = _get_health_expectations(self.config)
++        n_prefill, n_decode, count_desc, num_workers = _get_health_expectations(self.config, self.backend_processes)
+         logger.info("Waiting for server health (expecting %d health entries: %s)...", num_workers, count_desc)
+
+         hc = self.config.health_check

runners/launch_gb300-nv.sh

@@ -198,6 +198,14 @@ elif [[ "$IS_AGENTIC" == "1" ]]; then
         exit 1
     fi
     git cherry-pick --no-commit "$SRT_SLURM_MULTINODE_VLLM_PORT_SHA"
+
+    # Per-node DP launches one Dynamo generate endpoint per node-local process,
+    # not one per DP rank. Backport the health-count fix from
+    # ivanium/srt-slurm@ca0880138fa606130ae4acbb8d0afddfb84c69fa.
+    SRT_SLURM_PER_NODE_HEALTH_PATCH="$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-vllm-per-node-health.patch"
+    git apply --check "$SRT_SLURM_PER_NODE_HEALTH_PATCH"
+    git apply "$SRT_SLURM_PER_NODE_HEALTH_PATCH"
+
     mkdir -p recipes/vllm/deepseek-v4/agentic
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
         recipes/vllm/deepseek-v4/agentic