[Klaud Cold] minimaxm3-fp8-b200-vllm-mtp: day-zero MiniMax-M3 EAGLE3 (MTP) B200 recipe

.github/configs/nvidia-master.yaml

@@ -11845,6 +11845,40 @@ minimaxm3-fp8-b200-vllm:
       - { tp: 4, conc-start: 1, conc-end: 64 }
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 256 }
 
+# EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
+# minimaxm3-fp8-b200-vllm, pairing MiniMaxAI/MiniMax-M3-MXFP8 with the
+# Inferact/MiniMax-M3-EAGLE3 draft head (3 speculative tokens, drafter pinned
+# to FLASH_ATTN since the head is MHA and FlashInfer needs GQA/MQA at page
+# size 128). Search space mirrors the non-MTP entry trimmed at the
+# extreme-concurrency end, identical to the minimaxm3-fp8-b300-vllm-mtp
+# precedent: spec decode pays off at low/mid concurrency while acceptance
+# dilutes in big batches, and the draft weights + draft KV shave headroom.
+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: 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: 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: 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_fp8_b200_mtp.sh

@@ -0,0 +1,135 @@
+#!/usr/bin/env bash
+
+# MiniMax-M3 MXFP8 B200 single-node vLLM recipe with EAGLE3 speculative
+# decoding — the repo's spec-decoding=mtp variant of minimaxm3_fp8_b200.sh
+# (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3). Adds the
+# Inferact/MiniMax-M3-EAGLE3 draft head via --speculative-config with 3
+# speculative tokens. Everything else keeps the non-MTP serve shape:
+# --block-size 128 is mandatory (MSA sparse_block_size is 128; the default 16
+# misaligns sparse indexing), and --language-model-only skips the vision
+# encoder for the text-only benchmark. dp-attn=true maps to DP×EP (DEP);
+# ep>1 maps to TP+EP (TEP).
+#
+# The drafter is pinned to FLASH_ATTN: the EAGLE3 head is MHA, and FlashInfer
+# only supports page size 128 through its trtllm-gen kernel, which requires
+# GQA/MQA — engine init dies in FlashInferMetadataBuilder otherwise (the
+# failure hit on the B300 MTP canary). The target keeps its default
+# (FlashInfer) backend; FLASH_ATTN takes any multiple-of-16 block size, so
+# the mandatory 128 is fine for the draft.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    EP_SIZE \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3"
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+# launch_b200-dgxc.sh rewrites MODEL to the pre-downloaded
+# /lustre/fsw/gharunners/models/MiniMax-M3-MXFP8 path; only download the target
+# when handed a bare HF id (b200-cw / b200-nb runners). The EAGLE3 draft is
+# never pre-staged, so fetch it either way: next to the target weights when
+# MODEL is a local path (the gharunners tree is writable), into the HF cache
+# otherwise.
+if [[ "$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 "$DRAFT_MODEL_PATH"
+  fi
+fi
+
+SERVER_LOG=/workspace/server.log
+
+# 444 GB of MXFP8 weights off shared FS; engine startup can exceed the
+# default 600s readiness window.
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+
+if [ "${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
+
+# use 3 speculative tokens for all configs for now
+NUM_SPEC_TOKENS=3
+
+# Fixed-seq-len runs don't need graphs past the decode step's token count:
+# with spec decoding every running request contributes 1 + NUM_SPEC_TOKENS
+# tokens per step, so capture up to the next power of two >=
+# CONC * (1 + NUM_SPEC_TOKENS), capped at vLLM's 2048 ceiling.
+CAPTURE_SIZE=4
+while (( CAPTURE_SIZE < CONC * (1 + NUM_SPEC_TOKENS) )); do CAPTURE_SIZE=$((CAPTURE_SIZE * 2)); done
+(( CAPTURE_SIZE > 2048 )) && CAPTURE_SIZE=2048
+
+if [ "${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 \
+--max-cudagraph-capture-size $CAPTURE_SIZE \
+--max-num-batched-tokens "$((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 to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+pip install -q datasets pandas
+
+# Spec-decode acceptance rate degrades on raw random tokens; route prompts
+# through the chat template as the other MTP recipes do.
+run_benchmark_serving \
+    --model "$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

@@ -3696,3 +3696,13 @@
     - "H200 layouts: TP4 / TP8 (latency), TP4+EP4 / TP8+EP8 (TEP), TP8+EP8 dp-attn (DEP) across 1k1k and 8k1k — non-MTP search space trimmed at the extreme-concurrency end. H100 is TP8-only (no room below TP8 at ~56 GB weights/GPU); DEP omitted (KV-cache init fails at high conc, draft only tightens it)"
     - "Adds SPEC_SUFFIX to the three H100 launchers (cw, cr, dgxc-slurm) so spec-decoding=mtp routes to the _mtp script — they hardcoded _h100.sh and never gained the _mtp routing the H200 launchers have had since #392"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1739
+
+- config-keys:
+    - minimaxm3-fp8-b200-vllm-mtp
+  description:
+    - "Initial submission: MiniMax-M3 MXFP8 B200 vLLM benchmark with EAGLE3 speculative decoding (target: MiniMaxAI/MiniMax-M3-MXFP8, draft: Inferact/MiniMax-M3-EAGLE3, 3 speculative tokens) — spec-decoding=mtp variant of the day-zero B200 recipe"
+    - "Image: vllm/vllm-openai:minimax-m3 (same m3_release-branch build as the non-MTP entry)"
+    - "Serve shape follows minimaxm3-fp8-b200-vllm (--block-size 128, --language-model-only); cudagraph capture scaled to CONC * (1 + spec tokens); prompts routed through the chat template for realistic acceptance"
+    - "Drafter pinned to FLASH_ATTN via speculative-config attention_backend: the EAGLE3 head is MHA and FlashInfer only supports the mandatory page size 128 through its GQA-only trtllm-gen kernel"
+    - "Layouts: TP8 / TP4 (latency), TP8+EP8 / TP4+EP4 (TEP), TP8+EP8 dp-attn (DEP) across 1k1k and 8k1k — non-MTP search space trimmed at the extreme-concurrency end, mirroring the minimaxm3-fp8-b300-vllm-mtp search space"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1741