[NVIDIA] feat: MiniMax M3 Day 0 MTP (EAGLE3) support B300

.github/configs/nvidia-master.yaml

@@ -11750,3 +11750,31 @@ minimaxm3-fp8-b200-vllm:
       - { tp: 8, ep: 8, conc-start: 1, conc-end: 256 }
       - { tp: 4, conc-start: 1, conc-end: 64 }
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 256 }
+
+# MiniMax-M3 MXFP8 B300 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).
+# Latency-end focus (low/mid concurrency). No tp2 here: the EAGLE3 draft head
+# on top of ~222 GB/GPU MXFP8 weights leaves too little KV headroom at TP2.
+# Draft head served from the b300 /data/models tree (see launch_b300-nv.sh).
+minimaxm3-fp8-b300-vllm-mtp:
+  image: vllm/vllm-openai:minimax-m3
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: b300
+  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: 64, spec-decoding: mtp }

benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300_mtp.sh

@@ -0,0 +1,132 @@
+#!/usr/bin/env bash
+
+# MiniMax-M3 MXFP8 B300 single-node vLLM MTP (EAGLE3) variant of
+# minimaxm3_fp8_b300.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). Follows the b300 MODEL/MODEL_PATH split
+# (launch_b300-nv.sh serves from /data/models/<basename>). 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).
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    EP_SIZE \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE.
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+
+# EAGLE3 draft head. launch_b300-nv.sh stages models under /data/models (which
+# is bind-mounted), so the sibling of MODEL_PATH is visible in-container;
+# fall back to the HF id for stand-alone runs.
+DRAFT_MODEL="${DRAFT_MODEL_PATH:-}"
+if [[ -z "$DRAFT_MODEL" ]]; then
+    if [[ -d "${MODEL_PATH%/*}/MiniMax-M3-EAGLE3" ]]; then
+        DRAFT_MODEL="${MODEL_PATH%/*}/MiniMax-M3-EAGLE3"
+    else
+        DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3"
+    fi
+fi
+if [[ "$DRAFT_MODEL" != /* ]]; then hf download "$DRAFT_MODEL"; fi
+echo "EAGLE3 draft head: $DRAFT_MODEL"
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+SERVER_LOG=/workspace/server.log
+
+# 444 GB of MXFP8 weights + EAGLE3 draft head off shared FS; engine startup
+# can exceed the default 600s readiness window.
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+
+if [ "${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_PATH" --served-model-name "$MODEL" --host 0.0.0.0 --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

@@ -3656,3 +3656,13 @@
     - "Layouts: TP8 and TP4 (latency), TP4+EP4 / TP8+EP8 (TEP throughput), TP8+EP8 dp-attn (DEP) across 1k1k and 8k1k"
     - "Weights are not SRE-staged: launch_b200-dgxc.sh adds a MODEL_PATH case for /lustre/fsw/gharunners/models/MiniMax-M3-MXFP8; runners.yaml adds a b200-dgxc runner-type group"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1723
+
+- config-keys:
+    - minimaxm3-fp8-b300-vllm-mtp
+  description:
+    - "MiniMax-M3 MXFP8 B300 MTP (EAGLE3 speculative decoding) variant of minimaxm3-fp8-b300-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; no tp2 (EAGLE3 draft head + ~222 GB/GPU weights leaves too little KV headroom at TP2)"
+    - "EAGLE3 draft head served from the b300 /data/models tree (launch_b300-nv.sh)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1737