[NVIDIA] feat: MiniMax M3 Day 0 MTP (EAGLE3) support B200

.github/configs/nvidia-master.yaml

@@ -11850,3 +11850,31 @@ minimaxm3-fp8-b300-vllm-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: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 256, spec-decoding: mtp }
+
+# MiniMax-M3 MXFP8 B200 MTP (EAGLE3 spec decoding) — recipe spec_decoding
+# feature: Inferact/MiniMax-M3-EAGLE3 draft head, num_speculative_tokens=3,
+# attention_backend=FLASH_ATTN (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3).
+# Spec decoding accelerates the latency end (compute headroom to speculate), so
+# the sweep focuses on low/mid concurrency on the latency layouts. The draft
+# head is pre-staged beside the main weights on b200-dgxc (see launch_b200-dgxc.sh).
+minimaxm3-fp8-b200-vllm-mtp:
+  image: vllm/vllm-openai:minimax-m3
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: b200-dgxc
+  precision: fp8
+  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: 4, conc-start: 1, conc-end: 64, spec-decoding: mtp }
+      - { tp: 8, ep: 8, conc-start: 1, conc-end: 128, spec-decoding: mtp }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 1, conc-end: 32, spec-decoding: mtp }
+      - { tp: 4, conc-start: 1, conc-end: 32, spec-decoding: mtp }

benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh

@@ -0,0 +1,124 @@
+#!/usr/bin/env bash
+
+# MiniMax-M3 MXFP8 B200 single-node vLLM MTP (EAGLE3) variant of
+# minimaxm3_fp8_b200.sh (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3).
+# Adds the recipe's spec_decoding feature: EAGLE3 speculative decoding with the
+# Inferact/MiniMax-M3-EAGLE3 draft head (num_speculative_tokens=3,
+# attention_backend=FLASH_ATTN). EAGLE acceptance collapses on raw random
+# prompts, so the benchmark routes prompts through chat-formatted encoding via
+# --use-chat-template (required for all *_mtp.sh per AGENTS.md).
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    EP_SIZE \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+# launch_b200-dgxc.sh rewrites MODEL to the pre-downloaded
+# /lustre/fsw/gharunners/models/MiniMax-M3-MXFP8 path; only download when
+# handed a bare HF id (b200-cw / b200-nb runners).
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+# EAGLE3 draft head. launch_b200-dgxc.sh mounts it beside the main weights and
+# exports DRAFT_MODEL_PATH; default to the sibling of MODEL_PATH, then fall
+# back to the HF id (vLLM downloads it) for stand-alone runs.
+DRAFT_MODEL="${DRAFT_MODEL_PATH:-}"
+if [[ -z "$DRAFT_MODEL" ]]; then
+    if [[ "$MODEL" == /* && -d "${MODEL%/*}/MiniMax-M3-EAGLE3" ]]; then
+        DRAFT_MODEL="${MODEL%/*}/MiniMax-M3-EAGLE3"
+    else
+        DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3"
+    fi
+fi
+if [[ "$DRAFT_MODEL" != /* ]]; then hf download "$DRAFT_MODEL"; fi
+echo "EAGLE3 draft head: $DRAFT_MODEL"
+
+SERVER_LOG=/workspace/server.log
+
+# 444 GB of MXFP8 weights + EAGLE3 draft head off shared FS; engine startup
+# can exceed the default 600s readiness window.
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+
+if [ "${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
+
+# Fixed-seq-len runs don't need graphs past the request concurrency, but spec
+# decoding verifies CONC*(1+num_spec) tokens per decode step; capture up to the
+# next power of two >= that, capped at vLLM's 2048 ceiling.
+NUM_SPEC_TOKENS=3
+CAPTURE_SIZE=4
+TARGET=$(( CONC * (NUM_SPEC_TOKENS + 1) ))
+while (( CAPTURE_SIZE < TARGET )); do CAPTURE_SIZE=$((CAPTURE_SIZE * 2)); done
+(( CAPTURE_SIZE > 2048 )) && CAPTURE_SIZE=2048
+
+SPEC_CONFIG="{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\"}"
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+vllm serve $MODEL --port $PORT \
+$PARALLEL_ARGS \
+--gpu-memory-utilization 0.90 \
+--max-model-len $MAX_MODEL_LEN \
+--block-size 128 \
+--language-model-only \
+--speculative-config "$SPEC_CONFIG" \
+--max-cudagraph-capture-size $CAPTURE_SIZE \
+--max-num-batched-tokens "$((ISL * 2 ))" \
+--stream-interval 20 --no-enable-prefix-caching \
+--trust-remote-code > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+# --use-chat-template: EAGLE3 acceptance is trained against chat-formatted
+# inputs; benchmarking raw prompts silently regresses the acceptance rate.
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --use-chat-template
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x

perf-changelog.yaml

@@ -3677,3 +3677,13 @@
     - "Image: vllm/vllm-openai:minimax-m3 (dedicated day-zero image; M3 not in a stable release yet)."
     - "Sweeps TP4/TP8, TP+EP (TEP), and DP-attention+EP (DEP) per https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3; H100 is TP8-only (MXFP8 weights ~427 GB)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1731
+
+- config-keys:
+    - minimaxm3-fp8-b200-vllm-mtp
+  description:
+    - "MiniMax-M3 MXFP8 B200 MTP (EAGLE3 speculative decoding) variant of minimaxm3-fp8-b200-vllm"
+    - "Spec config (recipe spec_decoding feature): method eagle3, draft head Inferact/MiniMax-M3-EAGLE3, num_speculative_tokens=3, attention_backend FLASH_ATTN"
+    - "Image: vllm/vllm-openai:minimax-m3; --block-size 128, --language-model-only; benchmark uses --use-chat-template (EAGLE acceptance collapses on raw random prompts)"
+    - "Latency-end sweep (TP8, TP4, TP8+EP8) at low/mid concurrency where spec decoding has compute headroom to help"
+    - "EAGLE3 draft head pre-staged beside the main weights and bind-mounted on b200-dgxc (launch_b200-dgxc.sh)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1736

runners/launch_b200-dgxc.sh

@@ -68,6 +68,13 @@ elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
     # tree (root-owned); it lives in the sa-shared-writable gharunners tree.
     export MODEL_PATH="/lustre/fsw/gharunners/models/MiniMax-M3-MXFP8"
     export SRT_SLURM_MODEL_PREFIX="minimax-m3-mxfp8"
+    # EAGLE3 draft head for the MTP variant; pre-staged beside the main weights.
+    # Bind-mounted via EXTRA_MOUNTS below so the *_mtp.sh script sees it as the
+    # sibling of MODEL_PATH. Harmless for the non-MTP script (which ignores it).
+    if [[ -d "/lustre/fsw/gharunners/models/MiniMax-M3-EAGLE3" ]]; then
+        export DRAFT_MODEL_PATH="/lustre/fsw/gharunners/models/MiniMax-M3-EAGLE3"
+        EXTRA_MOUNTS="${EXTRA_MOUNTS:+$EXTRA_MOUNTS,}$DRAFT_MODEL_PATH:$DRAFT_MODEL_PATH"
+    fi
 else
     echo "Unsupported model prefix/precision: $MODEL_PREFIX/$PRECISION"
     echo "Available models under /lustre/fsw/models:"
@@ -448,7 +455,7 @@ else
 
     srun --jobid=$JOB_ID \
         --container-image=$SQUASH_FILE \
-        --container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$MODEL_PATH:$MODEL_PATH,$AIPERF_MMAP_CACHE_HOST_PATH:/aiperf_mmap_cache \
+        --container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$MODEL_PATH:$MODEL_PATH,$AIPERF_MMAP_CACHE_HOST_PATH:/aiperf_mmap_cache${EXTRA_MOUNTS:+,$EXTRA_MOUNTS} \
         --no-container-mount-home \
         --container-workdir=$CONTAINER_MOUNT_DIR \
         --no-container-entrypoint --export=ALL,PORT=8888,AIPERF_DATASET_MMAP_CACHE_DIR=/aiperf_mmap_cache \