Use MiniMax-M3 GB300 performance image and fix MNNVL workspace

.github/configs/nvidia-master.yaml

@@ -11835,7 +11835,7 @@ minimaxm3-fp8-b300-dynamo-vllm:
 # DEP8, DEP4. 4 GPU/node (GB300 NVL72). 4p3d (3 decode workers) skipped.
 # kv-cache-dtype=fp8 added. srun_options mem=0 required.
 minimaxm3-fp8-gb300-dynamo-vllm:
-  image: vllm/vllm-openai:nightly-aarch64
+  image: vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223
   model: MiniMaxAI/MiniMax-M3-MXFP8
   model-prefix: minimaxm3
   runner: gb300-nv

benchmarks/multi_node/srt-slurm-recipes/configs/minimax-m3-gb300-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-gb300-fp8/1k1k/disagg-gb300-1p1d-dep2-tep8-3n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-1p1d-dep2-tep8-fp8-1k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-dep2-tp4-marlin-2n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-1p1d-dep2-tp4-marlin-fp8-1k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p2d-dep2-dep4-3n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-1p2d-dep2-dep4-fp8-1k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p1d-dep2-dep8-3n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-2p1d-dep2-dep8-fp8-1k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p1d-dep2-tep8-3n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-2p1d-dep2-tep8-fp8-1k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p2d-dep2-tep8-5n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-2p2d-dep2-tep8-fp8-1k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-3p2d-dep2-tep8-6n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-3p2d-dep2-tep8-fp8-1k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p1d-dep2-tp4-marlin-2n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-1p1d-dep2-tp4-marlin-fp8-8k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p2d-dep2-dep8-5n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-1p2d-dep2-dep8-fp8-8k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-2p2d-dep2-dep8-5n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-2p2d-dep2-dep8-fp8-8k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-2p2d-dep2-tep8-5n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-2p2d-dep2-tep8-fp8-8k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-3p2d-dep2-dep8-6n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-3p2d-dep2-dep8-fp8-8k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-3p2d-dep2-tep8-6n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-3p2d-dep2-tep8-fp8-8k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-4p2d-dep2-dep8-6n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-4p2d-dep2-dep8-fp8-8k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-5p2d-dep2-tep8-7n.yaml

@@ -5,7 +5,7 @@ name: "minimax-m3-vllm-disagg-gb300-5p2d-dep2-tep8-fp8-8k1k"
 
 model:
   path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:nightly-aarch64"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
   precision: "fp8"
 
 dynamo:

perf-changelog.yaml

@@ -4099,3 +4099,11 @@
     - "Backport NVIDIA/srt-slurm#38 to sanitize Slurm node-IP discovery output on the pinned submission branch."
     - "Backport vllm-project/vllm#45879 so NIXL validates heterogeneous-TP KV block lengths using the GQA KV-head ratio."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1893
+
+- config-keys:
+    - minimaxm3-fp8-gb300-dynamo-vllm
+  description:
+    - "Update the GB300 MiniMax-M3 Dynamo-vLLM image to vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+    - "Use the dedicated ARM64 MiniMax-M3 performance image; benchmark settings unchanged"
+    - "Allocate FlashInfer MNNVL workspace for one-shot TP8 all-reduce during CUDA graph capture"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1888

runners/launch_gb300-nv.sh

@@ -106,6 +106,7 @@ export OSL="$OSL"
 echo "Cloning srt-slurm repository..."
 RUN_KEY=$(printf "%s" "${RESULT_FILENAME:-${RUNNER_NAME:-gb300-nv}}" | sha1sum | cut -c1-12)
 SRT_REPO_DIR="${GITHUB_WORKSPACE}/srt-slurm-${GITHUB_RUN_ID:-manual}-${GITHUB_RUN_ATTEMPT:-0}-${RUN_KEY}"
+SRTCTL_SETUP_SCRIPT=""
 rm -rf "$SRT_REPO_DIR"
 
 if [[ "$IS_AGENTIC" == "1" ]]; then
@@ -171,6 +172,10 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" ]]; then
     git checkout main
     mkdir -p recipes/vllm/minimax-m3-gb300-fp8
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8" recipes/vllm/minimax-m3-gb300-fp8
+    SRTCTL_SETUP_SCRIPT="minimax-m3-gb300-vllm-fixes.sh"
+    cp \
+        "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs/$SRTCTL_SETUP_SCRIPT" \
+        "configs/$SRTCTL_SETUP_SCRIPT"
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
@@ -274,12 +279,18 @@ sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_FILE"
 # seq-len recipes still resolve model.path to an NFS-visible location
 # where the precheck is a useful sanity guard, so keep enforcement on
 # for them.
-PREFLIGHT_FLAG=""
+SRTCTL_APPLY_ARGS=(
+    -f "$CONFIG_FILE"
+    --tags "gb300,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)"
+)
 if [[ "$IS_AGENTIC" == "1" ]]; then
-    PREFLIGHT_FLAG="--no-preflight"
+    SRTCTL_APPLY_ARGS+=(--no-preflight)
+fi
+if [[ -n "$SRTCTL_SETUP_SCRIPT" ]]; then
+    SRTCTL_APPLY_ARGS+=(--setup-script "$SRTCTL_SETUP_SCRIPT")
 fi
 
-SRTCTL_OUTPUT=$(srtctl apply $PREFLIGHT_FLAG -f "$CONFIG_FILE" --tags "gb300,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)" 2>&1)
+SRTCTL_OUTPUT=$(srtctl apply "${SRTCTL_APPLY_ARGS[@]}" 2>&1)
 echo "$SRTCTL_OUTPUT"
 
 JOB_ID=$(echo "$SRTCTL_OUTPUT" | grep -oP '✅ Job \K[0-9]+' || echo "$SRTCTL_OUTPUT" | grep -oP 'Job \K[0-9]+')