fixup

Author: Wei Zhao (Engrg-Hardware 1)

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb200-tp8-agentic.yaml

@@ -53,7 +53,9 @@ backend:
       global_segment_size: "150GB"
       local_buffer_size: "4GB"
       protocol: "rdma"
-      device_name: ""
+      # Pin to the IB compute fabric. device_name: "" auto-discovers RoCE/link-local
+      # NICs (mlx5_2/mlx5_5) that can not route cross-node and hang mooncake setup.
+      device_name: "mlx5_0,mlx5_1,mlx5_3,mlx5_4"
       mode: "embedded"
       enable_offload: false
   aggregated_environment:

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep12-agentic.yaml

@@ -60,7 +60,9 @@ backend:
       global_segment_size: "150GB"
       local_buffer_size: "4GB"
       protocol: "rdma"
-      device_name: ""
+      # Pin to the IB compute fabric. device_name: "" auto-discovers RoCE/link-local
+      # NICs (mlx5_2/mlx5_5) that can not route cross-node and hang mooncake setup.
+      device_name: "mlx5_0,mlx5_1,mlx5_3,mlx5_4"
       mode: "embedded"
       enable_offload: false
   prefill_environment:

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep8-agentic.yaml

@@ -60,7 +60,9 @@ backend:
       global_segment_size: "150GB"
       local_buffer_size: "4GB"
       protocol: "rdma"
-      device_name: ""
+      # Pin to the IB compute fabric. device_name: "" auto-discovers RoCE/link-local
+      # NICs (mlx5_2/mlx5_5) that can not route cross-node and hang mooncake setup.
+      device_name: "mlx5_0,mlx5_1,mlx5_3,mlx5_4"
       mode: "embedded"
       enable_offload: false
   prefill_environment:

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep12-agentic.yaml

@@ -60,7 +60,9 @@ backend:
       global_segment_size: "150GB"
       local_buffer_size: "4GB"
       protocol: "rdma"
-      device_name: ""
+      # Pin to the IB compute fabric. device_name: "" auto-discovers RoCE/link-local
+      # NICs (mlx5_2/mlx5_5) that can not route cross-node and hang mooncake setup.
+      device_name: "mlx5_0,mlx5_1,mlx5_3,mlx5_4"
       mode: "embedded"
       enable_offload: false
   prefill_environment:

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p1d-dep8-dep16-agentic.yaml

@@ -60,7 +60,9 @@ backend:
       global_segment_size: "150GB"
       local_buffer_size: "4GB"
       protocol: "rdma"
-      device_name: ""
+      # Pin to the IB compute fabric. device_name: "" auto-discovers RoCE/link-local
+      # NICs (mlx5_2/mlx5_5) that can not route cross-node and hang mooncake setup.
+      device_name: "mlx5_0,mlx5_1,mlx5_3,mlx5_4"
       mode: "embedded"
       enable_offload: false
   prefill_environment:

configs/nvidia-master.yaml

@@ -7663,35 +7663,41 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-2p1d-dep8-dep8:
           ep: 8
           dp-attn: true
 
-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
+# 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-1p1d-dep8-dep8:
   image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-c188b96

perf-changelog.yaml

@@ -4952,13 +4952,11 @@
   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 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]."
-    - "Aggregate TP8 uses sparse DSV4 attention, NUMA binding, max-num-seqs/CUDA graph size 32, max-num-batched-tokens 32768, and 0.95 GPU memory utilization."
+    - "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]."
     - "Image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-c188b96"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2260