feat: add MiniMax-M3 RTX 6000 Pro vLLM recipe / 新增 MiniMax-M3 RTX 6000 Pro vLLM 基准配置

benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_rtx6000pro.sh

@@ -0,0 +1,95 @@
+#!/usr/bin/env bash
+
+# MiniMax-M3 NVFP4 RTX PRO 6000 Blackwell single-node vLLM recipe.
+# This is the PCIe/SM120 counterpart to minimaxm3_fp4_b200.sh. It keeps
+# the ModelOpt NVFP4, FP8 KV-cache, and MSA block-size settings while using
+# NCCL collectives instead of the B200-tuned FlashInfer/TRT-LLM all-reduce.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    EP_SIZE \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+SERVER_LOG=/workspace/server.log
+GPU_MEM_UTIL="${GPU_MEM_UTIL:-0.90}"
+
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+export VLLM_FLOAT32_MATMUL_PRECISION=high
+
+if [ "${DP_ATTENTION}" = "true" ]; then
+    PARALLEL_ARGS=(
+        --tensor-parallel-size 1
+        --data-parallel-size "$TP"
+        --enable-expert-parallel
+    )
+elif [ "$EP_SIZE" -gt 1 ]; then
+    PARALLEL_ARGS=(
+        --tensor-parallel-size "$TP"
+        --enable-expert-parallel
+    )
+else
+    PARALLEL_ARGS=(--tensor-parallel-size "$TP")
+fi
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+start_gpu_monitor
+
+set -x
+vllm serve "$MODEL" --port "$PORT" \
+    "${PARALLEL_ARGS[@]}" \
+    --disable-custom-all-reduce \
+    --gpu-memory-utilization "$GPU_MEM_UTIL" \
+    --max-model-len "$MAX_MODEL_LEN" \
+    --kv-cache-dtype fp8 \
+    --block-size 128 \
+    --language-model-only \
+    --max-cudagraph-capture-size 2048 \
+    --max-num-batched-tokens "$((ISL * 2))" \
+    --stream-interval 20 \
+    --no-enable-prefix-caching \
+    --trust-remote-code > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x

configs/nvidia-master.yaml

@@ -7348,6 +7348,22 @@ minimaxm3-fp8-b300-vllm:
       - { tp: 4, ep: 4, dp-attn: true, conc-start: 64, conc-end: 128 }
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 512 }
 
+# MiniMax-M3 NVFP4 single-node vLLM bring-up on 8x RTX PRO 6000.
+minimaxm3-fp4-rtx6000pro-vllm:
+  image: vllm/vllm-openai:vllm-minimax-m3-perf-x86_64-13.0.1-8b00f41@sha256:6af4be7ae69a5f424de85b3d514ac79778bdcf4a9d05f93ec908a4095e0f1253
+  model: nvidia/MiniMax-M3-NVFP4
+  model-prefix: minimaxm3
+  runner: rtx6000pro-lat
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-list: [1] }
+
 # MiniMax-M3 NVFP4 (nvidia/MiniMax-M3-NVFP4) B300 single-node vLLM — FP4 variant
 # of minimaxm3-fp8-b300-vllm. MiniMax-M3 modelopt NVFP4 support (vllm-project/vllm
 # PR #46380) is baked into the perf container image, so no runtime patch is

configs/runners.yaml

@@ -164,6 +164,10 @@ labels:
     - gb300-nv_0
     - gb300-nv_1
     - gb300-nv_2
+  rtx6000pro:
+    - rtx6000pro-lat_00
+  rtx6000pro-lat:
+    - rtx6000pro-lat_00
   cluster:h100-cw:
     - h100-cw_00
     - h100-cw_01
@@ -251,6 +255,8 @@ labels:
     - gb300-nv_0
     - gb300-nv_1
     - gb300-nv_2
+  cluster:rtx6000pro-lat:
+    - rtx6000pro-lat_00
   cluster:mi300x-amds:
     - mi300x-amds_00
     - mi300x-amds_01
@@ -315,6 +321,9 @@ hardware:
   cluster:gb200-nv:
     available-cpu-dram-mib: 860_160
     gpus-per-node: 4
+  cluster:rtx6000pro-lat:
+    available-cpu-dram-mib: 1_500_000
+    gpus-per-node: 8
   cluster:mi300x-amds:
     available-cpu-dram-mib: 2_321_924
     gpus-per-node: 8

perf-changelog.yaml

@@ -5060,3 +5060,13 @@
     - "Re-pin VLLM_ROUTER_IMAGE to vllm/vllm-router:nightly-20260716-1fbcde7 (previous nightly-20260629-e667ebb was garbage-collected from Docker Hub)"
     - "Exclude known-bad nodes mia1-p01-g09,g14 from the disagg node pool"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2301
+
+- config-keys:
+    - minimaxm3-fp4-rtx6000pro-vllm
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Add MiniMax-M3 NVFP4 single-node vLLM benchmarking on the 8x RTX PRO 6000 Blackwell Latitude runner"
+    - "Start with a GPU-resident TP8 8k/1k concurrency-1 canary using FP8 KV cache, MSA block size 128, and NCCL collectives for the PCIe-only topology"
+    - "Pin the CUDA 13.0.1 MiniMax-M3 performance image by digest; it includes MiniMax-M3 ModelOpt NVFP4 support and SM120/121 FlashInfer FP4 kernels"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2306

runners/launch_rtx6000pro-lat.sh

@@ -0,0 +1,131 @@
+#!/usr/bin/bash
+set -euo pipefail
+
+# This runner executes directly on the single RTX PRO 6000 GPU node. Docker
+# therefore owns a separate image cache from the node's RKE2/containerd cache.
+HF_HUB_CACHE_MOUNT="${HF_HUB_CACHE_MOUNT:-/var/lib/inferencex/hf-hub-cache}"
+export HF_HUB_CACHE="${HF_HUB_CACHE:-/mnt/hf_hub_cache/}"
+PORT="${PORT:-8888}"
+
+: "${GITHUB_WORKSPACE:?GITHUB_WORKSPACE must be set}"
+: "${IMAGE:?IMAGE must be set}"
+: "${EXP_NAME:?EXP_NAME must be set}"
+: "${PRECISION:?PRECISION must be set}"
+
+mkdir -p "$HF_HUB_CACHE_MOUNT"
+
+export GPU_COUNT="${GPU_COUNT:-${TP:?TP must be set}}"
+if [[ ! "$GPU_COUNT" =~ ^[1-9][0-9]*$ ]]; then
+    echo "GPU_COUNT must be a positive integer, got: $GPU_COUNT" >&2
+    exit 1
+fi
+
+export CUDA_VISIBLE_DEVICES
+CUDA_VISIBLE_DEVICES="$(seq -s, 0 "$((GPU_COUNT - 1))")"
+
+# Some Slurm/enroot configs spell registry paths as nvcr.io#namespace/image.
+# Docker requires the normal slash form.
+DOCKER_IMAGE="${IMAGE//#//}"
+
+SPEC_SUFFIX=""
+if [[ "${SPEC_DECODING:-}" == "mtp" ]]; then
+    SPEC_SUFFIX="_mtp"
+fi
+
+export SCENARIO_SUBDIR="${SCENARIO_SUBDIR:-fixed_seq_len/}"
+SCENARIO_SUBDIR="${SCENARIO_SUBDIR#/}"
+SCENARIO_SUBDIR="${SCENARIO_SUBDIR%/}/"
+BENCH_SCRIPT="benchmarks/single_node/${SCENARIO_SUBDIR}${EXP_NAME%%_*}_${PRECISION}_rtx6000pro${SPEC_SUFFIX}.sh"
+
+if [[ ! -f "$GITHUB_WORKSPACE/$BENCH_SCRIPT" ]]; then
+    echo "Benchmark script not found: $GITHUB_WORKSPACE/$BENCH_SCRIPT" >&2
+    exit 1
+fi
+
+server_name="bmk-server-${RUNNER_NAME:-rtx6000pro-lat}"
+server_name="${server_name//[^a-zA-Z0-9_.-]/-}"
+
+cleanup() {
+    docker rm -f "$server_name" >/dev/null 2>&1 || true
+}
+trap cleanup EXIT
+
+# Clear a container left behind by a cancelled or interrupted workflow.
+cleanup
+
+docker run \
+    --rm \
+    --pull=missing \
+    --name="$server_name" \
+    --runtime=nvidia \
+    --gpus="$GPU_COUNT" \
+    --network=host \
+    --ipc=host \
+    --privileged \
+    --shm-size=32g \
+    --ulimit memlock=-1 \
+    --ulimit stack=67108864 \
+    --security-opt seccomp=unconfined \
+    --cap-add=SYS_PTRACE \
+    --volume "$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE" \
+    --volume "$GITHUB_WORKSPACE:/workspace/" \
+    --workdir=/workspace/ \
+    --env HF_TOKEN \
+    --env HF_HUB_CACHE \
+    --env MODEL \
+    --env MODEL_PREFIX \
+    --env MODEL_PATH \
+    --env TP \
+    --env PP_SIZE \
+    --env DCP_SIZE \
+    --env PCP_SIZE \
+    --env EP_SIZE \
+    --env DP_SIZE \
+    --env DP_ATTENTION \
+    --env GPU_COUNT \
+    --env CONC \
+    --env MAX_MODEL_LEN \
+    --env ISL \
+    --env OSL \
+    --env FRAMEWORK \
+    --env PRECISION \
+    --env DISAGG \
+    --env SPEC_DECODING \
+    --env NUM_SPEC_TOKENS \
+    --env RUN_EVAL \
+    --env EVAL_ONLY \
+    --env EVAL_LIMIT \
+    --env EVAL_MAX_MODEL_LEN \
+    --env RUNNER_TYPE \
+    --env RUNNER_NAME \
+    --env RESULT_FILENAME \
+    --env RESULT_DIR \
+    --env RANDOM_RANGE_RATIO \
+    --env GPU_MEM_UTIL \
+    --env AIPERF_FAILED_REQUEST_THRESHOLD \
+    --env KV_OFFLOADING \
+    --env KV_OFFLOAD_BACKEND \
+    --env KV_OFFLOAD_BACKEND_METADATA \
+    --env ROUTER_METADATA \
+    --env KV_P2P_TRANSFER \
+    --env TOTAL_CPU_DRAM_GB \
+    --env DURATION \
+    --env SCENARIO_TYPE \
+    --env SCENARIO_SUBDIR \
+    --env IS_AGENTIC \
+    --env SWEBENCH_GEN_MODE \
+    --env SWEBENCH_USE_MODAL \
+    --env MODAL_TOKEN_ID \
+    --env MODAL_TOKEN_SECRET \
+    --env PROFILE \
+    --env SGLANG_TORCH_PROFILER_DIR \
+    --env VLLM_TORCH_PROFILER_DIR \
+    --env VLLM_RPC_TIMEOUT \
+    --env PYTHONDONTWRITEBYTECODE \
+    --env PYTHONPYCACHEPREFIX=/tmp/pycache/ \
+    --env PORT="$PORT" \
+    --env CUDA_DEVICE_ORDER=PCI_BUS_ID \
+    --env CUDA_VISIBLE_DEVICES \
+    --entrypoint=/bin/bash \
+    "$DOCKER_IMAGE" \
+    "$BENCH_SCRIPT"