[Klaud Cold] minimaxm3 H200+H100 MTP: day-zero MiniMax-M3 EAGLE3 (MTP) recipes

.github/configs/nvidia-master.yaml

@@ -11878,3 +11878,64 @@ 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 }
+
+# EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
+# minimaxm3-fp8-h200-vllm (PR #1731), 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, per the dsv4-fp4-b300-vllm-mtp / minimaxm3 b300-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-h200-vllm-mtp:
+  image: vllm/vllm-openai:minimax-m3
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: h200
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 4, conc-start: 4, conc-end: 64, spec-decoding: mtp }
+      - { tp: 4, ep: 4, conc-start: 128, conc-end: 256, spec-decoding: mtp }
+      - { tp: 8, conc-start: 4, conc-end: 128, spec-decoding: mtp }
+      - { tp: 8, ep: 8, conc-start: 256, 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: 4, conc-start: 4, conc-end: 32, spec-decoding: mtp }
+      - { tp: 8, conc-start: 4, conc-end: 128, spec-decoding: mtp }
+      - { tp: 8, ep: 8, conc-start: 256, conc-end: 256, spec-decoding: mtp }
+      - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 256, spec-decoding: mtp }
+
+# EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
+# minimaxm3-fp8-h100-vllm (PR #1731). Same TP8-only layout (H100 80 GB has no
+# room below TP8 once ~56 GB of weights land per GPU) plus the
+# Inferact/MiniMax-M3-EAGLE3 draft head (3 speculative tokens, FLASH_ATTN
+# drafter). DEP stays omitted as on the non-MTP entry — KV-cache init already
+# failed at high conc, and the draft head + draft KV only tighten it further.
+minimaxm3-fp8-h100-vllm-mtp:
+  image: vllm/vllm-openai:minimax-m3
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: h100
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 128, spec-decoding: mtp }
+      - { tp: 8, ep: 8, conc-start: 256, conc-end: 256, spec-decoding: mtp }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: mtp }
+      - { tp: 8, ep: 8, conc-start: 128, conc-end: 256, spec-decoding: mtp }

benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_h100_mtp.sh

@@ -0,0 +1,157 @@
+#!/usr/bin/env bash
+
+# MiniMax-M3 MXFP8 H100 single-node vLLM recipe with EAGLE3 speculative
+# decoding — the repo's spec-decoding=mtp variant of minimaxm3_fp8_h100.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 benchmark
+# is text-only so --language-model-only frees the vision encoder's VRAM, and
+# the MXFP8 MoE runs through vLLM's Hopper-compatible backends. H100 is
+# TP8-only (~56 GB of weights per 80 GB GPU below TP8 leaves no room).
+#
+# 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 (hit on
+# the B300 MTP canary). 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 $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+# The shared HF cache lives on a network FS; concurrent day-zero downloads of
+# the same ~444 GB checkpoint from sibling nodes hit huggingface_hub's
+# WeakFileLock "[Errno 116] Stale file handle" race. Retry the download (it
+# resumes), then serve with HF_HUB_OFFLINE=1 so vllm's snapshot_download does
+# a lock-free local-cache read instead of re-contending the lock files. The
+# EAGLE3 draft is fetched the same way so the offline serve finds it cached.
+SERVE_OFFLINE=()
+if [[ "$MODEL" != /* ]]; then
+  for attempt in 1 2 3 4 5; do
+    hf download "$MODEL" && break
+    if [ "$attempt" = 5 ]; then echo "hf download failed after $attempt attempts" >&2; exit 1; fi
+    echo "hf download attempt $attempt failed; retrying in 60s" >&2
+    sleep 60
+  done
+  for attempt in 1 2 3 4 5; do
+    hf download "$DRAFT_MODEL" && break
+    if [ "$attempt" = 5 ]; then echo "hf download failed after $attempt attempts" >&2; exit 1; fi
+    echo "hf download attempt $attempt failed; retrying in 60s" >&2
+    sleep 60
+  done
+  SERVE_OFFLINE=(env HF_HUB_OFFLINE=1)
+fi
+
+SERVER_LOG=/workspace/server.log
+
+export PYTHONNOUSERSITE=1
+# ~444 GB of MXFP8 weights off shared FS; engine startup can exceed the
+# default 600s readiness window.
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+
+if [ "${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=$((CAPTURE_SIZE * 2)); done
+(( CAPTURE_SIZE > 2048 )) && CAPTURE_SIZE=2048
+
+# H100 DEP is weights-bound: every DP rank replicates the ~20 GB
+# BF16-dequantized attention/dense/embedding weights next to its ~52 GB
+# expert shard, and at gmu 0.90 KV-cache init fails outright at high conc
+# (sweep 27441767143, conc 256: "No available memory for the cache blocks").
+# Claw back headroom: higher gpu-memory-utilization, and decode graphs
+# capped at 2x the per-rank batch share (spec-token-scaled) instead of the
+# full CONC bound.
+GMU=0.90
+if [ "${DP_ATTENTION}" = "true" ]; then
+  GMU=0.94
+  PER_RANK_BOUND=$(( 2 * ((CONC + TP - 1) / TP) * (1 + NUM_SPEC_TOKENS) ))
+  CAPTURE_SIZE=4
+  while (( CAPTURE_SIZE < PER_RANK_BOUND )); do CAPTURE_SIZE=$((CAPTURE_SIZE * 2)); done
+  (( CAPTURE_SIZE > 2048 )) && CAPTURE_SIZE=2048
+fi
+
+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
+"${SERVE_OFFLINE[@]}" vllm serve $MODEL --port $PORT \
+$PARALLEL_ARGS \
+--gpu-memory-utilization $GMU \
+--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\", \"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

benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_h200_mtp.sh

@@ -0,0 +1,140 @@
+#!/usr/bin/env bash
+
+# MiniMax-M3 MXFP8 H200 single-node vLLM recipe with EAGLE3 speculative
+# decoding — the repo's spec-decoding=mtp variant of minimaxm3_fp8_h200.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 benchmark
+# is text-only so --language-model-only frees the vision encoder's VRAM, and
+# the MXFP8 MoE runs through vLLM's Hopper-compatible backends.
+#
+# 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 (hit on
+# the B300 MTP canary). 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 $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+# The shared HF cache lives on a network FS; concurrent day-zero downloads of
+# the same ~444 GB checkpoint from sibling nodes hit huggingface_hub's
+# WeakFileLock "[Errno 116] Stale file handle" race. Retry the download (it
+# resumes), then serve with HF_HUB_OFFLINE=1 so vllm's snapshot_download does
+# a lock-free local-cache read instead of re-contending the lock files. The
+# EAGLE3 draft is fetched the same way so the offline serve finds it cached.
+SERVE_OFFLINE=()
+if [[ "$MODEL" != /* ]]; then
+  for attempt in 1 2 3 4 5; do
+    hf download "$MODEL" && break
+    if [ "$attempt" = 5 ]; then echo "hf download failed after $attempt attempts" >&2; exit 1; fi
+    echo "hf download attempt $attempt failed; retrying in 60s" >&2
+    sleep 60
+  done
+  for attempt in 1 2 3 4 5; do
+    hf download "$DRAFT_MODEL" && break
+    if [ "$attempt" = 5 ]; then echo "hf download failed after $attempt attempts" >&2; exit 1; fi
+    echo "hf download attempt $attempt failed; retrying in 60s" >&2
+    sleep 60
+  done
+  SERVE_OFFLINE=(env HF_HUB_OFFLINE=1)
+fi
+
+SERVER_LOG=/workspace/server.log
+
+export PYTHONNOUSERSITE=1
+# ~444 GB of MXFP8 weights off shared FS; engine startup can exceed the
+# default 600s readiness window.
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+
+if [ "${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="$EVAL_MAX_MODEL_LEN"
+fi
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+"${SERVE_OFFLINE[@]}" 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\", \"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

@@ -3684,3 +3684,15 @@
     - "Add MiniMax-M2.5 NVFP4 B200 TensorRT-LLM single-node benchmark (1k1k and 8k1k)"
     - "Image: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc18"
   pr-link: https://github.com/NVIDIA/InferenceMAX/pull/1722
+
+- config-keys:
+    - minimaxm3-fp8-h200-vllm-mtp
+    - minimaxm3-fp8-h100-vllm-mtp
+  description:
+    - "Initial submission: MiniMax-M3 MXFP8 H200 + H100 vLLM benchmarks with EAGLE3 speculative decoding (target: MiniMaxAI/MiniMax-M3-MXFP8, draft: Inferact/MiniMax-M3-EAGLE3, 3 speculative tokens) — spec-decoding=mtp variants of the day-zero H200/H100 recipes (PR #1731)"
+    - "Image: vllm/vllm-openai:minimax-m3 (same m3_release-branch build as the non-MTP entries)"
+    - "Serve shape follows the non-MTP H200/H100 scripts (--block-size 128, --language-model-only, HF_HUB_OFFLINE serve to dodge the shared-FS download-lock race); 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"
+    - "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

runners/launch_h100-cr.sh

@@ -5,6 +5,10 @@ PORT=8888
 
 server_name="bmk-server"
 
+# Route spec-decoding=mtp configs to the _mtp benchmark script (parity with
+# the h200 launchers, which have carried SPEC_SUFFIX since #392).
+SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+
 set -x
 docker run --rm --network=host --name=$server_name \
 --runtime=nvidia --gpus=all --ipc=host --privileged --shm-size=16g --ulimit memlock=-1 --ulimit stack=67108864 \
@@ -15,4 +19,4 @@ docker run --rm --network=host --name=$server_name \
 -e PYTHONPYCACHEPREFIX=/tmp/pycache/ -e TORCH_CUDA_ARCH_LIST="9.0" -e CUDA_DEVICE_ORDER=PCI_BUS_ID -e CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" \
 --entrypoint=/bin/bash \
 $IMAGE \
-benchmarks/single_node/${SCENARIO_SUBDIR}"${EXP_NAME%%_*}_${PRECISION}_h100.sh"
+benchmarks/single_node/${SCENARIO_SUBDIR}"${EXP_NAME%%_*}_${PRECISION}_h100${SPEC_SUFFIX}.sh"