From e2e69c4fdfd6c1da7008b89c96ff8ebdec17180f Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Sat, 18 Jul 2026 11:47:43 -0400
Subject: [PATCH 01/11] feat(minimaxm3): add B200 spread-worker CI configs

---
 .../8k1k/1p1d-dep2-tp4-c1-c16-8k1k.yaml       |  62 +++++++++
 .../b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml | 127 ++++++++++++++++++
 .../8k1k/2p2d-dep2-tp4-c128-8k1k.yaml         | 127 ++++++++++++++++++
 .../8k1k/3p2d-dep2-tp4-c256-8k1k.yaml         | 127 ++++++++++++++++++
 .../8k1k/3p2d-dep2-tp4-c512-8k1k.yaml         | 127 ++++++++++++++++++
 .../8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml        | 127 ++++++++++++++++++
 configs/nvidia-master.yaml                    |  95 +++++++++++++
 perf-changelog.yaml                           |   8 ++
 runners/launch_b200-dgxc.sh                   |  10 +-
 9 files changed, 808 insertions(+), 2 deletions(-)
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p1d-dep2-tp4-c1-c16-8k1k.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p1d-dep2-tp4-c1-c16-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p1d-dep2-tp4-c1-c16-8k1k.yaml
new file mode 100644
index 0000000000..cd954e5124
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p1d-dep2-tp4-c1-c16-8k1k.yaml
@@ -0,0 +1,62 @@
+name: "minimax-m3-vllm-disagg-b200-1p1d-fp4-dep2-tp4-8k1k"
+model:
+  path: "nvidia/MiniMax-M3-NVFP4"
+  container: "vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
+  precision: "fp4"
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 2
+  gpus_per_decode: 4
+dynamo: {install: true, version: 1.3.0.dev20260710}
+frontend: {type: dynamo, enable_multiple_frontends: false}
+backend:
+  type: vllm
+  connector: null
+  allow_prefill_decode_colocation: true
+  prefill_environment: {VLLM_FLOAT32_MATMUL_PRECISION: high, UCX_TLS: "cuda_ipc,cuda_copy,rc"}
+  decode_environment: {VLLM_FLOAT32_MATMUL_PRECISION: high, UCX_TLS: "cuda_ipc,cuda_copy,rc"}
+  vllm_config:
+    prefill:
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      tensor-parallel-size: 1
+      data-parallel-size: 2
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      kv-cache-dtype: fp8
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+    decode:
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      tensor-parallel-size: 4
+      enable-expert-parallel: false
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      kv-cache-dtype: fp8
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      stream-interval: 32
+      max-num-seqs: 1024
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+health_check: {max_attempts: 360, interval_seconds: 10}
+benchmark: {type: "sa-bench", isl: 8192, osl: 1024, random_range_ratio: 0.8, concurrencies: "1x4x8x16", req_rate: "inf"}
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
new file mode 100644
index 0000000000..f326344be7
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
@@ -0,0 +1,127 @@
+name: "minimax-m3-vllm-disagg-b200-1p2d-fp4-dep2-tp4-c64-8k1k"
+model:
+  path: "nvidia/MiniMax-M3-NVFP4"
+  container: "vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
+  precision: "fp4"
+resources:
+  gpu_type: "b200"
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 2
+  decode_nodes: 2
+  decode_workers: 2
+  spread_workers: true
+  gpus_per_decode: 4
+  gpus_per_node: 8
+dynamo: {install: true, version: 1.3.0.dev20260710}
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    dyn-chat-processor: vllm
+    reasoning-parser: minimax_m3
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    VLLM_USE_DEEP_GEMM: '1'
+    VLLM_SKIP_P2P_CHECK: '1'
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: '1'
+    NVIDIA_GDRCOPY: '1'
+    PYTHONUNBUFFERED: '1'
+    VLLM_LOG_STATS_INTERVAL: '1'
+    NVSHMEM_IB_ENABLE_IBGDA: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '1'
+    NCCL_TIMEOUT: '1800'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
+    VLLM_USE_FLASHINFER_MOE_FP4: '1'
+    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
+    VLLM_USE_NCCL_SYMM_MEM: '1'
+    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
+    VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: cuda_ipc,cuda_copy,rc
+  decode_environment:
+    VLLM_USE_DEEP_GEMM: '1'
+    VLLM_SKIP_P2P_CHECK: '1'
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: '1'
+    NVIDIA_GDRCOPY: '1'
+    PYTHONUNBUFFERED: '1'
+    VLLM_LOG_STATS_INTERVAL: '1'
+    NVSHMEM_IB_ENABLE_IBGDA: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '1'
+    NCCL_TIMEOUT: '1800'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
+    VLLM_USE_FLASHINFER_MOE_FP4: '1'
+    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
+    VLLM_USE_NCCL_SYMM_MEM: '1'
+    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
+    VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
+    VLLM_MOE_DP_CHUNK_SIZE: '384'
+    VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD: '8192'
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: cuda_ipc,cuda_copy,rc
+  vllm_config:
+    prefill:
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      data-parallel-size: 2
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      max-model-len: 9472
+      max-num-seqs: 16
+      enforce-eager: true
+      gpu-memory-utilization: 0.9
+      max-num-batched-tokens: 16384
+      no-enable-chunked-prefill: true
+      kv-cache-dtype: fp8
+      async-scheduling: true
+      no-enable-prefix-caching: true
+      trust-remote-code: true
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      block-size: 128
+      language-model-only: true
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+    decode:
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-expert-parallel: false
+      max-model-len: 9472
+      max-num-seqs: 32
+      gpu-memory-utilization: 0.9
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","custom_ops":["+rms_norm"],"pass_config":{}}'
+      kv-cache-dtype: fp8
+      all2all-backend: deepep_low_latency
+      async-scheduling: true
+      stream-interval: 32
+      enable-dbo: true
+      dbo-decode-token-threshold: 32
+      no-enable-prefix-caching: true
+      trust-remote-code: true
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      block-size: 128
+      language-model-only: true
+health_check: {max_attempts: 360, interval_seconds: 10}
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "64"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
new file mode 100644
index 0000000000..6abfca3447
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
@@ -0,0 +1,127 @@
+name: "minimax-m3-vllm-disagg-b200-2p2d-fp4-dep2-tp4-c128-8k1k"
+model:
+  path: "nvidia/MiniMax-M3-NVFP4"
+  container: "vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
+  precision: "fp4"
+resources:
+  gpu_type: "b200"
+  prefill_nodes: 1
+  prefill_workers: 2
+  gpus_per_prefill: 2
+  decode_nodes: 2
+  decode_workers: 2
+  spread_workers: true
+  gpus_per_decode: 4
+  gpus_per_node: 8
+dynamo: {install: true, version: 1.3.0.dev20260710}
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    dyn-chat-processor: vllm
+    reasoning-parser: minimax_m3
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    VLLM_USE_DEEP_GEMM: '1'
+    VLLM_SKIP_P2P_CHECK: '1'
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: '1'
+    NVIDIA_GDRCOPY: '1'
+    PYTHONUNBUFFERED: '1'
+    VLLM_LOG_STATS_INTERVAL: '1'
+    NVSHMEM_IB_ENABLE_IBGDA: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '1'
+    NCCL_TIMEOUT: '1800'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
+    VLLM_USE_FLASHINFER_MOE_FP4: '1'
+    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
+    VLLM_USE_NCCL_SYMM_MEM: '1'
+    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
+    VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: cuda_ipc,cuda_copy,rc
+  decode_environment:
+    VLLM_USE_DEEP_GEMM: '1'
+    VLLM_SKIP_P2P_CHECK: '1'
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: '1'
+    NVIDIA_GDRCOPY: '1'
+    PYTHONUNBUFFERED: '1'
+    VLLM_LOG_STATS_INTERVAL: '1'
+    NVSHMEM_IB_ENABLE_IBGDA: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '1'
+    NCCL_TIMEOUT: '1800'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
+    VLLM_USE_FLASHINFER_MOE_FP4: '1'
+    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
+    VLLM_USE_NCCL_SYMM_MEM: '1'
+    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
+    VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
+    VLLM_MOE_DP_CHUNK_SIZE: '384'
+    VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD: '8192'
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: cuda_ipc,cuda_copy,rc
+  vllm_config:
+    prefill:
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      data-parallel-size: 2
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      max-model-len: 9472
+      max-num-seqs: 16
+      enforce-eager: true
+      gpu-memory-utilization: 0.9
+      max-num-batched-tokens: 16384
+      no-enable-chunked-prefill: true
+      kv-cache-dtype: fp8
+      async-scheduling: true
+      no-enable-prefix-caching: true
+      trust-remote-code: true
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      block-size: 128
+      language-model-only: true
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+    decode:
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-expert-parallel: false
+      max-model-len: 9472
+      max-num-seqs: 64
+      gpu-memory-utilization: 0.9
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","custom_ops":["+rms_norm"],"pass_config":{}}'
+      kv-cache-dtype: fp8
+      all2all-backend: deepep_low_latency
+      async-scheduling: true
+      stream-interval: 32
+      enable-dbo: true
+      dbo-decode-token-threshold: 32
+      no-enable-prefix-caching: true
+      trust-remote-code: true
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      block-size: 128
+      language-model-only: true
+health_check: {max_attempts: 360, interval_seconds: 10}
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "128"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
new file mode 100644
index 0000000000..a0d8ff18c9
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
@@ -0,0 +1,127 @@
+name: "minimax-m3-vllm-disagg-b200-3p2d-fp4-dep2-tp4-c256-8k1k"
+model:
+  path: "nvidia/MiniMax-M3-NVFP4"
+  container: "vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
+  precision: "fp4"
+resources:
+  gpu_type: "b200"
+  prefill_nodes: 1
+  prefill_workers: 3
+  gpus_per_prefill: 2
+  decode_nodes: 2
+  decode_workers: 2
+  spread_workers: true
+  gpus_per_decode: 4
+  gpus_per_node: 8
+dynamo: {install: true, version: 1.3.0.dev20260710}
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    dyn-chat-processor: vllm
+    reasoning-parser: minimax_m3
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    VLLM_USE_DEEP_GEMM: '1'
+    VLLM_SKIP_P2P_CHECK: '1'
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: '1'
+    NVIDIA_GDRCOPY: '1'
+    PYTHONUNBUFFERED: '1'
+    VLLM_LOG_STATS_INTERVAL: '1'
+    NVSHMEM_IB_ENABLE_IBGDA: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '1'
+    NCCL_TIMEOUT: '1800'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
+    VLLM_USE_FLASHINFER_MOE_FP4: '1'
+    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
+    VLLM_USE_NCCL_SYMM_MEM: '1'
+    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
+    VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: cuda_ipc,cuda_copy,rc
+  decode_environment:
+    VLLM_USE_DEEP_GEMM: '1'
+    VLLM_SKIP_P2P_CHECK: '1'
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: '1'
+    NVIDIA_GDRCOPY: '1'
+    PYTHONUNBUFFERED: '1'
+    VLLM_LOG_STATS_INTERVAL: '1'
+    NVSHMEM_IB_ENABLE_IBGDA: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '1'
+    NCCL_TIMEOUT: '1800'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
+    VLLM_USE_FLASHINFER_MOE_FP4: '1'
+    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
+    VLLM_USE_NCCL_SYMM_MEM: '1'
+    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
+    VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
+    VLLM_MOE_DP_CHUNK_SIZE: '384'
+    VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD: '8192'
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: cuda_ipc,cuda_copy,rc
+  vllm_config:
+    prefill:
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      data-parallel-size: 2
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      max-model-len: 9472
+      max-num-seqs: 16
+      enforce-eager: true
+      gpu-memory-utilization: 0.9
+      max-num-batched-tokens: 16384
+      no-enable-chunked-prefill: true
+      kv-cache-dtype: fp8
+      async-scheduling: true
+      no-enable-prefix-caching: true
+      trust-remote-code: true
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      block-size: 128
+      language-model-only: true
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+    decode:
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-expert-parallel: false
+      max-model-len: 9472
+      max-num-seqs: 128
+      gpu-memory-utilization: 0.9
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","custom_ops":["+rms_norm"],"pass_config":{}}'
+      kv-cache-dtype: fp8
+      all2all-backend: deepep_low_latency
+      async-scheduling: true
+      stream-interval: 32
+      enable-dbo: true
+      dbo-decode-token-threshold: 32
+      no-enable-prefix-caching: true
+      trust-remote-code: true
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      block-size: 128
+      language-model-only: true
+health_check: {max_attempts: 360, interval_seconds: 10}
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "256"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
new file mode 100644
index 0000000000..eb9705f776
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
@@ -0,0 +1,127 @@
+name: "minimax-m3-vllm-disagg-b200-3p2d-fp4-dep2-tp4-c512-8k1k"
+model:
+  path: "nvidia/MiniMax-M3-NVFP4"
+  container: "vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
+  precision: "fp4"
+resources:
+  gpu_type: "b200"
+  prefill_nodes: 1
+  prefill_workers: 3
+  gpus_per_prefill: 2
+  decode_nodes: 2
+  decode_workers: 2
+  spread_workers: true
+  gpus_per_decode: 4
+  gpus_per_node: 8
+dynamo: {install: true, version: 1.3.0.dev20260710}
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    dyn-chat-processor: vllm
+    reasoning-parser: minimax_m3
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    VLLM_USE_DEEP_GEMM: '1'
+    VLLM_SKIP_P2P_CHECK: '1'
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: '1'
+    NVIDIA_GDRCOPY: '1'
+    PYTHONUNBUFFERED: '1'
+    VLLM_LOG_STATS_INTERVAL: '1'
+    NVSHMEM_IB_ENABLE_IBGDA: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '1'
+    NCCL_TIMEOUT: '1800'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
+    VLLM_USE_FLASHINFER_MOE_FP4: '1'
+    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
+    VLLM_USE_NCCL_SYMM_MEM: '1'
+    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
+    VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: cuda_ipc,cuda_copy,rc
+  decode_environment:
+    VLLM_USE_DEEP_GEMM: '1'
+    VLLM_SKIP_P2P_CHECK: '1'
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: '1'
+    NVIDIA_GDRCOPY: '1'
+    PYTHONUNBUFFERED: '1'
+    VLLM_LOG_STATS_INTERVAL: '1'
+    NVSHMEM_IB_ENABLE_IBGDA: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '1'
+    NCCL_TIMEOUT: '1800'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
+    VLLM_USE_FLASHINFER_MOE_FP4: '1'
+    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
+    VLLM_USE_NCCL_SYMM_MEM: '1'
+    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
+    VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
+    VLLM_MOE_DP_CHUNK_SIZE: '384'
+    VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD: '8192'
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: cuda_ipc,cuda_copy,rc
+  vllm_config:
+    prefill:
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      data-parallel-size: 2
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      max-model-len: 9472
+      max-num-seqs: 16
+      enforce-eager: true
+      gpu-memory-utilization: 0.9
+      max-num-batched-tokens: 16384
+      no-enable-chunked-prefill: true
+      kv-cache-dtype: fp8
+      async-scheduling: true
+      no-enable-prefix-caching: true
+      trust-remote-code: true
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      block-size: 128
+      language-model-only: true
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+    decode:
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-expert-parallel: false
+      max-model-len: 9472
+      max-num-seqs: 256
+      gpu-memory-utilization: 0.9
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","custom_ops":["+rms_norm"],"pass_config":{}}'
+      kv-cache-dtype: fp8
+      all2all-backend: deepep_low_latency
+      async-scheduling: true
+      stream-interval: 32
+      enable-dbo: true
+      dbo-decode-token-threshold: 32
+      no-enable-prefix-caching: true
+      trust-remote-code: true
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      block-size: 128
+      language-model-only: true
+health_check: {max_attempts: 360, interval_seconds: 10}
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "512"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
new file mode 100644
index 0000000000..8f43fdf977
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
@@ -0,0 +1,127 @@
+name: "minimax-m3-vllm-disagg-b200-4p2d-fp4-dep2-tp4-c1024-8k1k"
+model:
+  path: "nvidia/MiniMax-M3-NVFP4"
+  container: "vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
+  precision: "fp4"
+resources:
+  gpu_type: "b200"
+  prefill_nodes: 2
+  prefill_workers: 4
+  gpus_per_prefill: 2
+  decode_nodes: 2
+  decode_workers: 2
+  spread_workers: true
+  gpus_per_decode: 4
+  gpus_per_node: 8
+dynamo: {install: true, version: 1.3.0.dev20260710}
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    dyn-chat-processor: vllm
+    reasoning-parser: minimax_m3
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    VLLM_USE_DEEP_GEMM: '1'
+    VLLM_SKIP_P2P_CHECK: '1'
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: '1'
+    NVIDIA_GDRCOPY: '1'
+    PYTHONUNBUFFERED: '1'
+    VLLM_LOG_STATS_INTERVAL: '1'
+    NVSHMEM_IB_ENABLE_IBGDA: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '1'
+    NCCL_TIMEOUT: '1800'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
+    VLLM_USE_FLASHINFER_MOE_FP4: '1'
+    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
+    VLLM_USE_NCCL_SYMM_MEM: '1'
+    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
+    VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: cuda_ipc,cuda_copy,rc
+  decode_environment:
+    VLLM_USE_DEEP_GEMM: '1'
+    VLLM_SKIP_P2P_CHECK: '1'
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: '1'
+    NVIDIA_GDRCOPY: '1'
+    PYTHONUNBUFFERED: '1'
+    VLLM_LOG_STATS_INTERVAL: '1'
+    NVSHMEM_IB_ENABLE_IBGDA: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '1'
+    NCCL_TIMEOUT: '1800'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
+    VLLM_USE_FLASHINFER_MOE_FP4: '1'
+    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
+    VLLM_USE_NCCL_SYMM_MEM: '1'
+    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
+    VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
+    VLLM_MOE_DP_CHUNK_SIZE: '384'
+    VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD: '8192'
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: cuda_ipc,cuda_copy,rc
+  vllm_config:
+    prefill:
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      data-parallel-size: 2
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      max-model-len: 9472
+      max-num-seqs: 16
+      enforce-eager: true
+      gpu-memory-utilization: 0.9
+      max-num-batched-tokens: 16384
+      no-enable-chunked-prefill: true
+      kv-cache-dtype: fp8
+      async-scheduling: true
+      no-enable-prefix-caching: true
+      trust-remote-code: true
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      block-size: 128
+      language-model-only: true
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+    decode:
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-expert-parallel: false
+      max-model-len: 9472
+      max-num-seqs: 512
+      gpu-memory-utilization: 0.9
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","custom_ops":["+rms_norm"],"pass_config":{}}'
+      kv-cache-dtype: fp8
+      all2all-backend: deepep_low_latency
+      async-scheduling: true
+      stream-interval: 32
+      enable-dbo: true
+      dbo-decode-token-threshold: 32
+      no-enable-prefix-caching: true
+      trust-remote-code: true
+      served-model-name: nvidia/MiniMax-M3-NVFP4
+      no-enable-flashinfer-autotune: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      block-size: 128
+      language-model-only: true
+health_check: {max_attempts: 360, interval_seconds: 10}
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "1024"
+  req_rate: "inf"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 3aac1b8378..6fade36f6c 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -6445,6 +6445,101 @@ qwen3.5-fp8-h100-sglang-agentic:
       - { tp: 8, ep: 8, kv-offloading: none,    conc-list: [1, 2, 4, 8, 12, 14, 16] }
       - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [12, 14, 16, 20, 24, 28, 32, 42] }
 
+minimaxm3-fp4-b200-dynamo-vllm:
+  image: vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2
+  model: nvidia/MiniMax-M3-NVFP4
+  model-prefix: minimaxm3
+  runner: b200-multinode
+  precision: fp4
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.3.0.dev20260710" }
+  kv-p2p-transfer: nixl
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - conc-list: [1, 4, 8, 16]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/1p1d-dep2-tp4-c1-c16-8k1k.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [64]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [128]
+        prefill:
+          num-worker: 2
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [256]
+        prefill:
+          num-worker: 3
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [512]
+        prefill:
+          num-worker: 3
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [1024]
+        prefill:
+          num-worker: 4
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 4
+          ep: 1
+          dp-attn: false
+
 # MiniMax-M3 NVFP4 disagg sweep on the same B300 topology matrix as the MXFP8
 # baseline above. The image includes vLLM PR #46380, so no runtime patch is
 # needed.
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 005967fc36..0df24385f1 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4966,3 +4966,11 @@
     - "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:
+    - minimaxm3-fp4-b200-dynamo-vllm
+  description:
+    - "Add MiniMax-M3 NVFP4 B200 Dynamo-vLLM disaggregated benchmarks at 8k/1k."
+    - "Use configurations from concurrency 1 through 1024 with spread worker placement for CI transport validation."
+    - "Image: vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2; Dynamo: 1.3.0.dev20260710."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2230
diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
index 58051e6405..23d7a1aa3a 100644
--- a/runners/launch_b200-dgxc.sh
+++ b/runners/launch_b200-dgxc.sh
@@ -69,9 +69,8 @@ elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
     export MODEL_PATH="/lustre/fsw/gharunners/models/MiniMax-M3-MXFP8"
     export SRT_SLURM_MODEL_PREFIX="minimax-m3-mxfp8"
 elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" ]]; then
-    # NVFP4 checkpoint, pre-staged on the b200-dgxc scratch tree.
     export MODEL_PATH="/scratch/fsw/models/MiniMax-M3-NVFP4"
-    export SRT_SLURM_MODEL_PREFIX="minimax-m3-nvfp4"
+    export SRT_SLURM_MODEL_PREFIX="nvidia/MiniMax-M3-NVFP4"
 else
     echo "Unsupported model prefix/precision: $MODEL_PREFIX/$PRECISION"
     echo "Available models under /lustre/fsw/models:"
@@ -113,6 +112,13 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         git checkout aflowers/vllm-gb200-v0.20.0
         mkdir -p recipes/vllm/deepseek-v4
         cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
+    elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" ]]; then
+        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+        cd "$SRT_REPO_DIR" || exit 1
+        mkdir -p recipes/vllm/minimax-m3/b200-fp4
+        cp -rT \
+            "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4" \
+            recipes/vllm/minimax-m3/b200-fp4
     elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5" && $PRECISION == "fp8" ]]; then
         git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
         cd "$SRT_REPO_DIR" || exit 1

From 39266bb803f5a71c8de1d9a3bda677a485111ae2 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Sun, 19 Jul 2026 17:48:37 -0400
Subject: [PATCH 02/11] fix(minimaxm3): reserve nodes for spread workers

---
 .../vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml  | 2 +-
 .../vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml  | 2 +-
 .../vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml  | 2 +-
 .../vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml | 2 +-
 4 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
index 6abfca3447..11c0b1d978 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
@@ -5,7 +5,7 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 1
+  prefill_nodes: 2
   prefill_workers: 2
   gpus_per_prefill: 2
   decode_nodes: 2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
index a0d8ff18c9..dbdf8f1803 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
@@ -5,7 +5,7 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 1
+  prefill_nodes: 3
   prefill_workers: 3
   gpus_per_prefill: 2
   decode_nodes: 2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
index eb9705f776..59a4d0eca4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
@@ -5,7 +5,7 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 1
+  prefill_nodes: 3
   prefill_workers: 3
   gpus_per_prefill: 2
   decode_nodes: 2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
index 8f43fdf977..9e18dfedad 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
@@ -5,7 +5,7 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 2
+  prefill_nodes: 4
   prefill_workers: 4
   gpus_per_prefill: 2
   decode_nodes: 2

From c7d8fe24b50168bfe7d37a8bb5b631ca5c4ea6cd Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Mon, 20 Jul 2026 19:54:27 -0400
Subject: [PATCH 03/11] feat(minimaxm3): use selective worker placement

---
 .../minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml     | 3 ++-
 .../minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml    | 5 +++--
 .../minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml    | 5 +++--
 .../minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml    | 5 +++--
 .../minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml   | 5 +++--
 5 files changed, 14 insertions(+), 9 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
index f326344be7..307c1186f5 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
@@ -7,10 +7,11 @@ resources:
   gpu_type: "b200"
   prefill_nodes: 1
   prefill_workers: 1
+  prefill_workers_per_node: 1
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
-  spread_workers: true
+  spread_decode_workers: true
   gpus_per_decode: 4
   gpus_per_node: 8
 dynamo: {install: true, version: 1.3.0.dev20260710}
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
index 11c0b1d978..cfd337defc 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
@@ -5,12 +5,13 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 2
+  prefill_nodes: 1
   prefill_workers: 2
+  prefill_workers_per_node: 2
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
-  spread_workers: true
+  spread_decode_workers: true
   gpus_per_decode: 4
   gpus_per_node: 8
 dynamo: {install: true, version: 1.3.0.dev20260710}
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
index dbdf8f1803..51429fe52b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
@@ -5,12 +5,13 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 3
+  prefill_nodes: 1
   prefill_workers: 3
+  prefill_workers_per_node: 3
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
-  spread_workers: true
+  spread_decode_workers: true
   gpus_per_decode: 4
   gpus_per_node: 8
 dynamo: {install: true, version: 1.3.0.dev20260710}
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
index 59a4d0eca4..e3666df2f9 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
@@ -5,12 +5,13 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 3
+  prefill_nodes: 1
   prefill_workers: 3
+  prefill_workers_per_node: 3
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
-  spread_workers: true
+  spread_decode_workers: true
   gpus_per_decode: 4
   gpus_per_node: 8
 dynamo: {install: true, version: 1.3.0.dev20260710}
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
index 9e18dfedad..8fab07a8bf 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
@@ -5,12 +5,13 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 4
+  prefill_nodes: 1
   prefill_workers: 4
+  prefill_workers_per_node: 4
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
-  spread_workers: true
+  spread_decode_workers: true
   gpus_per_decode: 4
   gpus_per_node: 8
 dynamo: {install: true, version: 1.3.0.dev20260710}

From 2d0622f6b9638ba61577d3935a8ddb79fa2829e3 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Wed, 22 Jul 2026 10:22:18 -0400
Subject: [PATCH 04/11] new

---
 runners/launch_b200-dgxc.sh | 1 +
 1 file changed, 1 insertion(+)

diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
index 23d7a1aa3a..36115a9aeb 100644
--- a/runners/launch_b200-dgxc.sh
+++ b/runners/launch_b200-dgxc.sh
@@ -115,6 +115,7 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" ]]; then
         git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
         cd "$SRT_REPO_DIR" || exit 1
+        git checkout srt-slurm-placement
         mkdir -p recipes/vllm/minimax-m3/b200-fp4
         cp -rT \
             "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4" \

From 462a7fe61a7a58c13eb5fd3d7e28eea2536f3f7c Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Thu, 23 Jul 2026 00:55:35 -0400
Subject: [PATCH 05/11] update ucx configs

---
 .../vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml   | 2 --
 .../vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml  | 1 -
 .../vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml  | 1 -
 .../vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml  | 1 -
 .../vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml | 1 -
 5 files changed, 6 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
index 307c1186f5..7fd45e6bed 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
@@ -40,7 +40,6 @@ backend:
     VLLM_USE_FLASHINFER_MOE_FP4: '1'
     VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
     VLLM_USE_NCCL_SYMM_MEM: '1'
-    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
     VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
     VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
     VLLM_FLOAT32_MATMUL_PRECISION: high
@@ -61,7 +60,6 @@ backend:
     VLLM_USE_FLASHINFER_MOE_FP4: '1'
     VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
     VLLM_USE_NCCL_SYMM_MEM: '1'
-    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
     VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
     VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
     VLLM_MOE_DP_CHUNK_SIZE: '384'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
index cfd337defc..e8d1bdbf44 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
@@ -40,7 +40,6 @@ backend:
     VLLM_USE_FLASHINFER_MOE_FP4: '1'
     VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
     VLLM_USE_NCCL_SYMM_MEM: '1'
-    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
     VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
     VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
     VLLM_FLOAT32_MATMUL_PRECISION: high
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
index 51429fe52b..ebbc213cd4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
@@ -40,7 +40,6 @@ backend:
     VLLM_USE_FLASHINFER_MOE_FP4: '1'
     VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
     VLLM_USE_NCCL_SYMM_MEM: '1'
-    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
     VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
     VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
     VLLM_FLOAT32_MATMUL_PRECISION: high
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
index e3666df2f9..18e43705db 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
@@ -40,7 +40,6 @@ backend:
     VLLM_USE_FLASHINFER_MOE_FP4: '1'
     VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
     VLLM_USE_NCCL_SYMM_MEM: '1'
-    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
     VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
     VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
     VLLM_FLOAT32_MATMUL_PRECISION: high
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
index 8fab07a8bf..475f83b32d 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
@@ -40,7 +40,6 @@ backend:
     VLLM_USE_FLASHINFER_MOE_FP4: '1'
     VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: '0'
     VLLM_USE_NCCL_SYMM_MEM: '1'
-    UCX_IB_ROCE_REACHABILITY_MODE: local_subnet
     VLLM_NIXL_SIDE_CHANNEL_PORT: '5600'
     VLLM_NIXL_ABORT_REQUEST_TIMEOUT: '300'
     VLLM_FLOAT32_MATMUL_PRECISION: high

From f97abea75419a2c0a1b59e17093d1a7bb6726955 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Fri, 24 Jul 2026 09:49:36 -0400
Subject: [PATCH 06/11] remove srt-slurm-placement branch checkout; rename
 spread_decode_workers to spread_workers

---
 .../vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml   | 2 +-
 .../vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml  | 2 +-
 .../vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml  | 2 +-
 .../vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml  | 2 +-
 .../vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml | 2 +-
 runners/launch_b200-dgxc.sh                                     | 1 -
 6 files changed, 5 insertions(+), 6 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
index ae7bfc3c1d..1b669f1d14 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
@@ -11,7 +11,7 @@ resources:
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
-  spread_decode_workers: true
+  spread_workers: true
   gpus_per_decode: 4
   gpus_per_node: 8
 dynamo: {install: true, version: 1.3.0.dev20260710}
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
index 0efa7ab05a..284e08f530 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
@@ -11,7 +11,7 @@ resources:
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
-  spread_decode_workers: true
+  spread_workers: true
   gpus_per_decode: 4
   gpus_per_node: 8
 dynamo: {install: true, version: 1.3.0.dev20260710}
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
index 999841514c..3e84d517cf 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
@@ -11,7 +11,7 @@ resources:
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
-  spread_decode_workers: true
+  spread_workers: true
   gpus_per_decode: 4
   gpus_per_node: 8
 dynamo: {install: true, version: 1.3.0.dev20260710}
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
index fc1a4533c2..0664dd300f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
@@ -11,7 +11,7 @@ resources:
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
-  spread_decode_workers: true
+  spread_workers: true
   gpus_per_decode: 4
   gpus_per_node: 8
 dynamo: {install: true, version: 1.3.0.dev20260710}
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
index 630ef60eaa..751ab247fd 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
@@ -11,7 +11,7 @@ resources:
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
-  spread_decode_workers: true
+  spread_workers: true
   gpus_per_decode: 4
   gpus_per_node: 8
 dynamo: {install: true, version: 1.3.0.dev20260710}
diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
index 26b5d67fb4..93ccec8913 100644
--- a/runners/launch_b200-dgxc.sh
+++ b/runners/launch_b200-dgxc.sh
@@ -115,7 +115,6 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" ]]; then
         git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
         cd "$SRT_REPO_DIR" || exit 1
-        git checkout srt-slurm-placement
         mkdir -p recipes/vllm/minimax-m3/b200-fp4
         cp -rT \
             "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4" \

From 244fa9a90bb40af1d10566a13b162c96f08aaae7 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Fri, 24 Jul 2026 09:57:27 -0400
Subject: [PATCH 07/11] remove prefill_workers_per_node from minimax-m3
 b200-fp4 disagg configs

---
 .../vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml    | 1 -
 .../vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml   | 1 -
 .../vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml   | 1 -
 .../vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml   | 1 -
 .../vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml  | 1 -
 5 files changed, 5 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
index 1b669f1d14..7545f73b29 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/1p2d-dep2-tp4-c64-8k1k.yaml
@@ -7,7 +7,6 @@ resources:
   gpu_type: "b200"
   prefill_nodes: 1
   prefill_workers: 1
-  prefill_workers_per_node: 1
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
index 284e08f530..5a0a933172 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
@@ -7,7 +7,6 @@ resources:
   gpu_type: "b200"
   prefill_nodes: 1
   prefill_workers: 2
-  prefill_workers_per_node: 2
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
index 3e84d517cf..85564bb236 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
@@ -7,7 +7,6 @@ resources:
   gpu_type: "b200"
   prefill_nodes: 1
   prefill_workers: 3
-  prefill_workers_per_node: 3
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
index 0664dd300f..7ece5952c0 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
@@ -7,7 +7,6 @@ resources:
   gpu_type: "b200"
   prefill_nodes: 1
   prefill_workers: 3
-  prefill_workers_per_node: 3
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
index 751ab247fd..a922d6c587 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
@@ -7,7 +7,6 @@ resources:
   gpu_type: "b200"
   prefill_nodes: 1
   prefill_workers: 4
-  prefill_workers_per_node: 4
   gpus_per_prefill: 2
   decode_nodes: 2
   decode_workers: 2

From e427fb963fd9e7f4a248f957ad8d61bc9322e7cb Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Fri, 24 Jul 2026 10:29:38 -0400
Subject: [PATCH 08/11] fix prefill_nodes to match prefill_workers for
 spread_workers topology

---
 .../vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml  | 2 +-
 .../vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml  | 2 +-
 .../vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml  | 2 +-
 .../vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml | 2 +-
 4 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
index 5a0a933172..4614abeae0 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/2p2d-dep2-tp4-c128-8k1k.yaml
@@ -5,7 +5,7 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 1
+  prefill_nodes: 2
   prefill_workers: 2
   gpus_per_prefill: 2
   decode_nodes: 2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
index 85564bb236..8cfb3eb300 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c256-8k1k.yaml
@@ -5,7 +5,7 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 1
+  prefill_nodes: 3
   prefill_workers: 3
   gpus_per_prefill: 2
   decode_nodes: 2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
index 7ece5952c0..bf9f0d7e59 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/3p2d-dep2-tp4-c512-8k1k.yaml
@@ -5,7 +5,7 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 1
+  prefill_nodes: 3
   prefill_workers: 3
   gpus_per_prefill: 2
   decode_nodes: 2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
index a922d6c587..ff366b74eb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/4p2d-dep2-tp4-c1024-8k1k.yaml
@@ -5,7 +5,7 @@ model:
   precision: "fp4"
 resources:
   gpu_type: "b200"
-  prefill_nodes: 1
+  prefill_nodes: 4
   prefill_workers: 4
   gpus_per_prefill: 2
   decode_nodes: 2

From 2f2559a483b19c8fb5b21278412ddf8dfeadd611 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Fri, 24 Jul 2026 14:41:31 -0400
Subject: [PATCH 09/11] revert b300-fp4 mtp configs to original (no image/UCX
 changes needed)

---
 .../b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml        | 6 +++---
 .../b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml        | 6 +++---
 .../b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml        | 6 +++---
 .../b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml        | 6 +++---
 .../b300-fp4/8k1k/mtp/2p1d-dep2-dep4-eagle3-8k1k.yaml       | 6 +++---
 .../b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml        | 6 +++---
 6 files changed, 18 insertions(+), 18 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml
index dcd545353a..0890ace22f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml
@@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-1p1d-dep2-tp4-fp4-8k1k-eagle3"
 
 model:
   path: "nvidia/MiniMax-M3-NVFP4"
-  container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9"
+  container: "vllm/vllm-openai:nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
   precision: fp4
 
 resources:
@@ -32,12 +32,12 @@ backend:
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml
index 734ee42d32..19cdeb0793 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml
@@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-1p2d-dep2-tp4-fp4-8k1k-eagle3"
 
 model:
   path: "nvidia/MiniMax-M3-NVFP4"
-  container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9"
+  container: "vllm/vllm-openai:nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
   precision: fp4
 
 resources:
@@ -32,12 +32,12 @@ backend:
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml
index 958686c617..858576055b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml
@@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-1p4d-dep2-tp4-fp4-8k1k-eagle3"
 
 model:
   path: "nvidia/MiniMax-M3-NVFP4"
-  container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9"
+  container: "vllm/vllm-openai:nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
   precision: fp4
 
 resources:
@@ -32,12 +32,12 @@ backend:
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml
index b2f13b532f..b4da6d26e4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml
@@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-1p6d-dep2-tp4-fp4-8k1k-eagle3"
 
 model:
   path: "nvidia/MiniMax-M3-NVFP4"
-  container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9"
+  container: "vllm/vllm-openai:nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
   precision: fp4
 
 resources:
@@ -32,12 +32,12 @@ backend:
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p1d-dep2-dep4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p1d-dep2-dep4-eagle3-8k1k.yaml
index 44c9ab677b..5943109612 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p1d-dep2-dep4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p1d-dep2-dep4-eagle3-8k1k.yaml
@@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-2p1d-dep2-dep4-fp4-8k1k-eagle3"
 
 model:
   path: "nvidia/MiniMax-M3-NVFP4"
-  container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9"
+  container: "vllm/vllm-openai:nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
   precision: fp4
 
 resources:
@@ -32,12 +32,12 @@ backend:
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml
index 1f33f1ba92..73aa7c161c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml
@@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-2p3d-dep2-tp4-fp4-8k1k-eagle3"
 
 model:
   path: "nvidia/MiniMax-M3-NVFP4"
-  container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9"
+  container: "vllm/vllm-openai:nightly-8e981630c9336233ca9de91452f68918bddbc4e2"
   precision: fp4
 
 resources:
@@ -32,12 +32,12 @@ backend:
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
     VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm
-    UCX_TCP_AF_PRIO: "inet"
+    UCX_TLS: cuda_copy,cuda_ipc,rc
 
   vllm_config:
     prefill:

From 6ebf4aca15d6e5a1240ebb1e90fda8feb88904e9 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Fri, 24 Jul 2026 14:43:32 -0400
Subject: [PATCH 10/11] update minimaxm3-fp4-b200-dynamo-vllm image in
 nvidia-master.yaml

---
 configs/nvidia-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 8777c586ad..8cfc8ad2df 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -6664,7 +6664,7 @@ qwen3.5-fp8-h100-sglang-agentic:
       - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [12, 14, 16, 20, 24, 28, 32, 42] }
 
 minimaxm3-fp4-b200-dynamo-vllm:
-  image: vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2
+  image: vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9
   model: nvidia/MiniMax-M3-NVFP4
   model-prefix: minimaxm3
   runner: b200-multinode

From 3f245da10da506c645fc15a0f57833977da8f62b Mon Sep 17 00:00:00 2001
From: Ankur-singh <ankusingh@nvidia.com>
Date: Sat, 25 Jul 2026 12:38:55 -0700
Subject: [PATCH 11/11] perf-changelog: correct minimaxm3-fp4-b200-dynamo-vllm
 image string [skip-sweep]

The config and all six b200-fp4 recipes pin
vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9, but the
changelog entry still cited the pre-bump cu129-nightly-8e981630c9 tag.
Documentation-only fix; no benchmark or config behavior changes.
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index e501b2e863..b9890b8410 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5055,7 +5055,7 @@
   description:
     - "Add MiniMax-M3 NVFP4 B200 Dynamo-vLLM disaggregated benchmarks at 8k/1k."
     - "Use configurations from concurrency 1 through 1024 with spread worker placement for CI transport validation."
-    - "Image: vllm/vllm-openai:cu129-nightly-8e981630c9336233ca9de91452f68918bddbc4e2; Dynamo: 1.3.0.dev20260710."
+    - "Image: vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9; Dynamo: 1.3.0.dev20260710."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2230
 
 - config-keys: