MiniMax-M3 GB200 full sweep on the performance vLLM image

.github/configs/nvidia-master.yaml

@@ -12104,6 +12104,252 @@ qwen3.5-fp4-b200-trt:
       - { tp: 4, ep: 4, dp-attn: true, conc-list: [1024] }
       - { tp: 8, ep: 8, dp-attn: true, conc-list: [256, 512, 1024] }
 
+# MiniMax-M3 GB200 disagg sweep — adapted from NV B300 PR #1863.
+# All prefill DEP4 (TP1 DP4 EP, 4 GPU/worker). Decode: TP4+Marlin, TEP8,
+# DEP8, DEP4. 4 GPU/node (GB200 NVL72). 4p3d (3 decode workers) skipped.
+# FLASHINFER attention with FP8 KV cache, matching the validated GB300 sweep.
+minimaxm3-fp8-gb200-dynamo-vllm:
+  image: vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: gb200
+  precision: fp8
+  framework: dynamo-vllm
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      # 1p1d DEP4+TEP8, 3n: conc 4,16,64,128,4096
+      - conc-list: [4, 16, 64, 128, 4096]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/disagg-gb200-1p1d-dep4-tep8-3n.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: false
+
+      # 1p1d DEP4+TP4 Marlin, 2n: conc 1,4,8,16
+      - conc-list: [1, 4, 8, 16]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/disagg-gb200-1p1d-dep4-tp4-marlin-2n.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+
+      # 1p2d DEP4+DEP4, 3n: conc 2048
+      - conc-list: [2048]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/disagg-gb200-1p2d-dep4-dep4-3n.yaml"
+        decode:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+
+      # 2p1d DEP4+DEP8, 4n: conc 512,4096
+      - conc-list: [512, 4096]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/disagg-gb200-2p1d-dep4-dep8-4n.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+      # 2p1d DEP4+TEP8, 4n: conc 32
+      - conc-list: [32]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/disagg-gb200-2p1d-dep4-tep8-4n.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: false
+
+      # 2p2d DEP4+TEP8, 6n: conc 16
+      - conc-list: [16]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/disagg-gb200-2p2d-dep4-tep8-6n.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: false
+
+      # 3p2d DEP4+TEP8, 7n: conc 4
+      - conc-list: [4]
+        prefill:
+          num-worker: 3
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/disagg-gb200-3p2d-dep4-tep8-7n.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: false
+
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 1p1d DEP4+TP4 Marlin, 2n: conc 1,4,8,16
+      - conc-list: [1, 4, 8, 16]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/disagg-gb200-1p1d-dep4-tp4-marlin-2n.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+
+      # 1p2d DEP4+DEP8, 5n: conc 128
+      - conc-list: [128]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/disagg-gb200-1p2d-dep4-dep8-5n.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+      # 2p2d DEP4+DEP8, 6n: conc 256,512
+      - conc-list: [256, 512]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/disagg-gb200-2p2d-dep4-dep8-6n.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+      # 2p2d DEP4+TEP8, 6n: conc 16
+      - conc-list: [16]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/disagg-gb200-2p2d-dep4-tep8-6n.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: false
+
+      # 3p2d DEP4+DEP8, 7n: conc 512
+      - conc-list: [512]
+        prefill:
+          num-worker: 3
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/disagg-gb200-3p2d-dep4-dep8-7n.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+      # 3p2d DEP4+TEP8, 7n: conc 32
+      - conc-list: [32]
+        prefill:
+          num-worker: 3
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/disagg-gb200-3p2d-dep4-tep8-7n.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: false
+
+      # 4p2d DEP4+DEP8, 8n: conc 4096
+      - conc-list: [4096]
+        prefill:
+          num-worker: 4
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/disagg-gb200-4p2d-dep4-dep8-8n.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+      # 5p2d DEP4+TEP8, 9n: conc 4,64
+      - conc-list: [4, 64]
+        prefill:
+          num-worker: 5
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/disagg-gb200-5p2d-dep4-tep8-9n.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: false
+
 # MiniMax-M3 day-zero (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3).
 # 427B total / 26B active MoE with MSA sparse attention; MXFP8 checkpoint
 # (MiniMaxAI/MiniMax-M3-MXFP8, ~444 GB) quantized by NVIDIA — native MX tensor

benchmarks/multi_node/srt-slurm-recipes/configs/minimax-m3-gb200-vllm-fixes.sh

@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+python3 - <<'PYEOF'
+from importlib.util import find_spec
+from pathlib import Path
+
+spec = find_spec("vllm")
+if not spec or not spec.origin:
+    raise RuntimeError("vllm is not installed")
+root = Path(spec.origin).parent
+patches = {
+    root / "distributed/device_communicators/flashinfer_all_reduce.py": [
+        (
+            "            comm_backend=comm_backend,\n"
+            "            group=group,\n",
+            "            comm_backend=comm_backend,\n"
+            '            force_oneshot_support=backend == "mnnvl",\n'
+            "            group=group,\n",
+        ),
+    ],
+    root / "models/minimax_m3/nvidia/sparse_attention_msa.py": [
+        (
+            "            prefill_topk = topk[:, nd:num_tokens, :]\n",
+            "            prefill_topk = topk[:, nd:num_tokens, :].contiguous()\n",
+        ),
+    ],
+}
+for path, edits in patches.items():
+    source = path.read_text()
+    for old, new in edits:
+        if new in source:
+            continue
+        if source.count(old) != 1:
+            raise RuntimeError(f"missing or ambiguous patch anchor in {path}")
+        source = source.replace(old, new, 1)
+    path.write_text(source)
+PYEOF

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/disagg-gb200-1p1d-dep4-tep8-3n.yaml

@@ -0,0 +1,101 @@
+name: "minimax-m3-vllm-disagg-gb200-1p1d-dep4-tep8-fp8-1k1k"
+
+# 1P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 1D TEP8 (TP8+EP8)
+# 3 nodes (1P + 2D). Adapted from NV B300 PR #1863.
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 4
+  gpus_per_decode: 8
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_TLS: "cuda_copy,cuda_ipc,tcp"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_TLS: "cuda_copy,cuda_ipc,tcp"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      attention-backend: FLASHINFER
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      tensor-parallel-size: 8
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      attention-backend: FLASHINFER
+      stream-interval: 32
+      max-num-seqs: 4096
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 8192
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "4x16x64x128x4096"
+  req_rate: "inf"