Re-add tp8 config with patch

Wei Zhao (Engrg-Hardware 1) · ivanium · commit 40da599f97ba · 2026-07-19T08:40:52.000Z
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb200-tp8-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb200-tp8-agentic.yaml
@@ -18,7 +18,7 @@ slurm:
   time_limit: "8:00:00"
 
 health_check:
-  max_attempts: 2160
+  max_attempts: 720
   interval_seconds: 10
 
 resources:
@@ -63,7 +63,11 @@ backend:
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_RPC_TIMEOUT: "600000"
     TILELANG_CLEANUP_TEMP_FILES: "1"
-    VLLM_USE_NCCL_SYMM_MEM: "1"
+    # Off: the c188b96 container's CUDA-graph capture path enters vLLM's
+    # nccl_symm_mem_context without set_graph_pool_id, so capture asserts
+    # "graph_pool_id is not set under graph capture". Re-enable once this recipe
+    # moves to a container built from a vLLM with the cuda_graph.py fix.
+    VLLM_USE_NCCL_SYMM_MEM: "0"
     TORCH_SYMMMEM: "NVSHMEM"
     NCCL_CUMEM_ENABLE: "1"
     NCCL_MNNVL_ENABLE: "1"
@@ -97,7 +101,11 @@ backend:
       tensor-parallel-size: 8
       pipeline-parallel-size: 1
       enable-cumem-allocator: true
-      numa-bind: true
+      # numa-bind disabled: srt-slurm's cpus-per-task/cgroup already constrains
+      # CPU affinity, so vLLM's auto-NUMA detection bails and --numa-bind hard-
+      # errors ("could not detect the GPU-to-NUMA topology"). Re-enable only with
+      # an explicit --numa-bind-nodes mapping for this cluster.
+      # numa-bind: true
       attention-config: '{"backend": "FLASHINFER_MLA_SPARSE_DSV4", "use_prefill_query_quantization": true, "use_fp4_indexer_cache": true}'
       max-model-len: 1048576
       # AgentX session fan-out can exceed the configured live-session
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep12-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep12-agentic.yaml
@@ -18,7 +18,7 @@ slurm:
   time_limit: "8:00:00"
 
 health_check:
-  max_attempts: 2160
+  max_attempts: 720
   interval_seconds: 10
 
 resources:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep8-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep8-agentic.yaml
@@ -18,7 +18,7 @@ slurm:
   time_limit: "8:00:00"
 
 health_check:
-  max_attempts: 2160
+  max_attempts: 720
   interval_seconds: 10
 
 resources:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep12-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep12-agentic.yaml
@@ -18,7 +18,7 @@ slurm:
   time_limit: "8:00:00"
 
 health_check:
-  max_attempts: 2160
+  max_attempts: 720
   interval_seconds: 10
 
 resources:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-agentic.yaml
@@ -20,7 +20,7 @@ slurm:
   time_limit: "8:00:00"
 
 health_check:
-  max_attempts: 1440
+  max_attempts: 720
   interval_seconds: 10
 
 resources:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p1d-dep8-dep16-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p1d-dep8-dep16-agentic.yaml
@@ -18,7 +18,7 @@ slurm:
   time_limit: "8:00:00"
 
 health_check:
-  max_attempts: 2160
+  max_attempts: 720
   interval_seconds: 10
 
 resources:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p2d-tep8-tp8-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p2d-tep8-tp8-agentic.yaml
@@ -21,7 +21,7 @@ slurm:
   time_limit: "8:00:00"
 
 health_check:
-  max_attempts: 1440
+  max_attempts: 720
   interval_seconds: 10
 
 resources:
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
@@ -7663,41 +7663,35 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-2p1d-dep8-dep8:
           ep: 8
           dp-attn: true
 
-# TEMPORARILY DISABLED: the TP8 aggregate is a single vLLM engine with TP=8
-# spanning 2 GB200 nodes (4 GPUs each). It hits a vLLM shm-broadcast port race
-# on the follower node -- the 4 remote-node TP workers all call get_open_port()
-# from the same VLLM_PORT and collide (EADDRINUSE) with no retry, crashing the
-# engine. Re-enable once that vLLM race is fixed. The disagg configs below use
-# per-node DEP (not cross-node single-engine TP8) and are unaffected.
-# dsv4-fp4-gb200-dynamo-vllm-agentic-agg:
-#   image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-c188b96
-#   model: deepseek-ai/DeepSeek-V4-Pro
-#   model-prefix: dsv4
-#   runner: cluster:gb200-nv
-#   precision: fp4
-#   framework: dynamo-vllm
-#   router: { name: dynamo-router, version: "1.3.0.dev20260618" }
-#   multinode: true
-#   disagg: false
-#   scenarios:
-#     agentic-coding:
-#     - search-space:
-#       - spec-decoding: none
-#         conc-list: [1, 4, 8, 16]
-#         prefill:
-#           num-worker: 1
-#           tp: 8
-#           ep: 1
-#           dp-attn: false
-#           additional-settings:
-#           - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/agg-gb200-tp8-agentic.yaml"
-#         # The aggregate worker also performs decode; keep the decode worker
-#         # count at zero so result aggregation counts eight GPUs only once.
-#         decode:
-#           num-worker: 0
-#           tp: 8
-#           ep: 1
-#           dp-attn: false
+dsv4-fp4-gb200-dynamo-vllm-agentic-agg:
+  image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-c188b96
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:gb200-nv
+  precision: fp4
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.3.0.dev20260618" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - spec-decoding: none
+        conc-list: [1, 4, 8, 16]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/agg-gb200-tp8-agentic.yaml"
+        # The aggregate worker also performs decode; keep the decode worker
+        # count at zero so result aggregation counts eight GPUs only once.
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 1
+          dp-attn: false
 
 dsv4-fp4-gb200-dynamo-vllm-agentic-1p1d-dep8-dep8:
   image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-c188b96
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
@@ -4952,11 +4952,12 @@
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2256
 
 - config-keys:
+    - dsv4-fp4-gb200-dynamo-vllm-agentic-agg
     - dsv4-fp4-gb200-dynamo-vllm-agentic-1p1d-dep8-dep8
     - dsv4-fp4-gb200-dynamo-vllm-agentic-1p1d-dep8-dep12
     - dsv4-fp4-gb200-dynamo-vllm-agentic-2p1d-dep8-dep12
     - dsv4-fp4-gb200-dynamo-vllm-agentic-3p1d-dep8-dep16
   description:
-    - "Add GB200 Dynamo-vLLM AgentX disaggregated topologies: 1P/1D DEP8/DEP8 at [64,128,256,320], 1P/1D DEP8/DEP12 at [128,256,320], 2P/1D DEP8/DEP12 at [480,640,768], and 3P/1D DEP8/DEP16 at [640,800,960,1280]."
+    - "Add GB200 Dynamo-vLLM AgentX aggregate TP8 at conc [1,4,8,16] and disaggregated topologies: 1P/1D DEP8/DEP8 at [64,128,256,320], 1P/1D DEP8/DEP12 at [128,256,320], 2P/1D DEP8/DEP12 at [480,640,768], and 3P/1D DEP8/DEP16 at [640,800,960,1280]."
     - "Image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-c188b96"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2260
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
@@ -319,6 +319,14 @@ if [[ "$IS_AGENTIC" == "1" ]]; then
     fi
     git cherry-pick --no-commit "$SRT_SLURM_PER_NODE_DP_SHA" || exit 1
 
+    # Pin VLLM_PORT only for single-GPU processes. Multi-GPU worker processes
+    # (multi-node tensor parallel, e.g. the TP8 aggregate spanning two 4-GPU
+    # nodes) run vLLM's internal multiproc executor whose same-node subprocesses
+    # otherwise all read one VLLM_PORT and race to bind the shm-broadcast port
+    # (EADDRINUSE crash). Not yet in an srt-slurm release, so apply it here as a
+    # patch until it lands upstream.
+    git apply "$GITHUB_WORKSPACE/runners/patches/srt-slurm-vllm-port-single-gpu.patch" || exit 1
+
     # ai-dynamo/dynamo#11303 is merged into ai-dynamo/dynamo main, so the
     # recipe-pinned dynamo hash resolves against upstream directly -- no
     # esmeetu/dynamo fork redirect of srt-slurm's schema.py needed anymore.
diff --git a/runners/patches/srt-slurm-vllm-port-single-gpu.patch b/runners/patches/srt-slurm-vllm-port-single-gpu.patch
@@ -0,0 +1,36 @@
+diff --git a/src/srtctl/backends/vllm.py b/src/srtctl/backends/vllm.py
+index 0484d70..f3a95ca 100644
+--- a/src/srtctl/backends/vllm.py
++++ b/src/srtctl/backends/vllm.py
+@@ -284,7 +284,8 @@ class VLLMProtocol:
+           multi-node NIXL handshake)
+         - VLLM_PORT: private base for vLLM's get_open_port() scans, unique per
+           process so co-located workers don't race for the same rendezvous port
+-          (see the notes on VLLM_PORT_BASE in srtctl.ports)
++          (see the notes on VLLM_PORT_BASE in srtctl.ports). Set ONLY for
++          single-GPU processes -- see below.
+         """
+         from srtctl.core.slurm import get_hostname_ip
+ 
+@@ -299,8 +300,19 @@ class VLLMProtocol:
+         # 4xGB200 nodes: each prefill endpoint is DEP2 (uses 2 of the 4 GPUs), so
+         # two endpoints share one physical node and would otherwise scan
+         # overlapping get_open_port() ranges.
+-        proc_index = max(process.sys_port - DYN_SYSTEM_PORT_BASE, 0)
+-        env["VLLM_PORT"] = str(VLLM_PORT_BASE + proc_index * VLLM_PORT_STRIDE)
++        #
++        # Only pin it for single-GPU processes (the per-GPU DP layout). A
++        # multi-GPU process runs vLLM's internal multiproc executor, whose N
++        # same-node worker subprocesses all read this one VLLM_PORT; in a
++        # multi-node TP group the remote-node subprocesses then race to bind the
++        # shm-broadcast port from that shared base and crash with EADDRINUSE
++        # (e.g. the TP8 aggregate spanning two 4-GPU nodes). Leaving VLLM_PORT
++        # unset lets those subprocesses fall back to OS-assigned ephemeral ports,
++        # which are unique per bind. Single-node multi-GPU TP never hit this
++        # (same-node reader -> IPC, no TCP bind), so this is safe there too.
++        if len(process.gpu_indices) == 1:
++            proc_index = max(process.sys_port - DYN_SYSTEM_PORT_BASE, 0)
++            env["VLLM_PORT"] = str(VLLM_PORT_BASE + proc_index * VLLM_PORT_STRIDE)
+         return env
+ 
+     def get_mooncake_worker_env(self, infra_node_ip: str, local_hostname: str) -> dict[str, str]:
