Add MiniMax-M3 NVFP4 B200 single-node vLLM benchmark (EAGLE3 spec decode)

.github/configs/nvidia-master.yaml

@@ -12369,6 +12369,44 @@ minimaxm3-fp8-b200-vllm-mtp:
       - { tp: 4, ep: 4, dp-attn: true, conc-start: 64, conc-end: 128, spec-decoding: mtp }
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 256, spec-decoding: mtp }
 
+# EAGLE3 speculative-decoding (spec-decoding: mtp) variant of MiniMax-M3 NVFP4
+# (nvidia/MiniMax-M3-NVFP4) B200 single-node vLLM, pairing the target with the
+# Inferact/MiniMax-M3-EAGLE3 draft head (3 speculative tokens). Runs on the
+# b200-dgxc cluster. MiniMax-M3 modelopt NVFP4 support (vllm-project/vllm
+# PR #46380) is baked into the perf container image, so no runtime patch is
+# needed; prompts are routed through the chat template. Target weights are
+# pre-staged at /scratch/fsw/models/MiniMax-M3-NVFP4 (launch_b200-dgxc.sh
+# resolves MODEL_PATH for minimaxm3-fp4); the EAGLE3 draft is fetched next to
+# the target weights.
+minimaxm3-fp4-b200-vllm-mtp:
+  image: vllm/vllm-openai:vllm-minimax-m3-perf-x86_64-13.0.1-8b00f41
+  model: nvidia/MiniMax-M3-NVFP4
+  model-prefix: minimaxm3
+  runner: b200-dgxc
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 1, conc-end: 64, spec-decoding: mtp }
+      - { tp: 8, ep: 8, conc-start: 1, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, conc-start: 1, conc-end: 64, spec-decoding: mtp }
+      - { tp: 4, ep: 4, conc-start: 64, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, ep: 4, dp-attn: true, conc-start: 128, conc-end: 512, spec-decoding: mtp }
+      - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 512, spec-decoding: mtp }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 1, conc-end: 64, spec-decoding: mtp }
+      - { tp: 8, ep: 8, conc-start: 1, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, conc-start: 1, conc-end: 64, spec-decoding: mtp }
+      - { tp: 4, ep: 4, conc-start: 64, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, ep: 4, dp-attn: true, conc-start: 64, conc-end: 128, spec-decoding: mtp }
+      - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 256, spec-decoding: mtp }
+
 # EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
 # minimaxm3-fp8-b300-vllm, pairing MiniMaxAI/MiniMax-M3-MXFP8 with the
 # Inferact/MiniMax-M3-EAGLE3 draft head (3 speculative tokens).

benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200_mtp.sh

@@ -0,0 +1,107 @@
+#!/usr/bin/env bash
+
+# MiniMax-M3 NVFP4 B200 single-node vLLM recipe with EAGLE3 speculative
+# decoding — same shape as minimaxm3_fp8_b200_mtp.sh but uses the
+# nvidia/MiniMax-M3-NVFP4 checkpoint. MiniMax-M3 modelopt NVFP4 support
+# (vllm-project/vllm PR #46380) is baked into the perf container image, so no
+# runtime patch is needed.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    EP_SIZE \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3"
+
+# launch_b200-dgxc.sh rewrites MODEL to the pre-downloaded path; only download
+# the target when handed a bare HF id (b200-cw / b200-nb runners). The EAGLE3
+# draft is never pre-staged, so fetch it either way: next to the target weights
+# when MODEL is a local path, into the HF cache otherwise.
+if [[ "$MODEL" != /* ]]; then
+  hf download "$MODEL"
+  hf download "$DRAFT_MODEL"
+  DRAFT_MODEL_PATH="$DRAFT_MODEL"
+else
+  DRAFT_MODEL_PATH="$(dirname "$MODEL")/${DRAFT_MODEL##*/}"
+  if [[ ! -d "$DRAFT_MODEL_PATH" || -z "$(ls -A "$DRAFT_MODEL_PATH" 2>/dev/null)" ]]; then
+    hf download "$DRAFT_MODEL" --local-dir "$DRAFT_MODEL_PATH"
+  fi
+fi
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+SERVER_LOG=/workspace/server.log
+
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+export VLLM_FLOAT32_MATMUL_PRECISION=high
+
+if [ "${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
+
+# use 3 speculative tokens for all configs for now
+NUM_SPEC_TOKENS=3
+
+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 \
+--gpu-memory-utilization 0.90 \
+--max-model-len $MAX_MODEL_LEN \
+--block-size 128 \
+--language-model-only \
+--max-cudagraph-capture-size 2048 \
+--max-num-batched-tokens "$((ISL * 2 ))" \
+--speculative-config "{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL_PATH\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\"}" \
+--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 "$SERVER_PID"
+
+pip install -q datasets pandas
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --use-chat-template
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x

perf-changelog.yaml

@@ -4196,3 +4196,12 @@
   description:
     - "Initial submission: MiniMax-M3 MXFP4 disagg (prefill/decode) on MI355X with vLLM over the MoRI-IO KV connector (8k/1k)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1914
+
+- config-keys:
+    - minimaxm3-fp4-b200-vllm-mtp
+  description:
+    - "Add MiniMax-M3 NVFP4 (nvidia/MiniMax-M3-NVFP4) B200 single-node aggregated vLLM benchmark with EAGLE3 speculative decoding (spec-decoding: mtp, 3 draft tokens via Inferact/MiniMax-M3-EAGLE3), runner b200-dgxc"
+    - "Image vllm/vllm-openai:vllm-minimax-m3-perf-x86_64-13.0.1-8b00f41 (bakes in MiniMax-M3 modelopt NVFP4 support, vllm-project/vllm PR #46380; no runtime patch needed); prompts routed through the chat template"
+    - "Target weights pre-staged at /scratch/fsw/models/MiniMax-M3-NVFP4 (added minimaxm3-fp4 MODEL_PATH branch to launch_b200-dgxc.sh); EAGLE3 draft fetched next to the target weights; --block-size 128 (MSA), --language-model-only"
+    - "Sweeps tp 4/8 with and without EP and dp-attn at 1k1k and 8k1k, conc 1-512"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1933

runners/launch_b200-dgxc.sh

@@ -68,6 +68,10 @@ elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
     # tree (root-owned); it lives in the sa-shared-writable gharunners tree.
     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"
 else
     echo "Unsupported model prefix/precision: $MODEL_PREFIX/$PRECISION"
     echo "Available models under /lustre/fsw/models:"