From b5520374eff7abfa79a8e8e34e6766b2bfb522de Mon Sep 17 00:00:00 2001 From: Cam Quilici Date: Fri, 12 Jun 2026 14:49:48 -0500 Subject: [PATCH 1/4] [NVIDIA] feat: MiniMax M3 Day 0 support B300 MXFP8 single-node vLLM sweep (TP/TEP/DEP, incl. tp2-ep2) for MiniMax-M3 on B300. --block-size 128 (MSA sparse attention), --language-model-only for text-only throughput, dedicated vllm/vllm-openai:minimax-m3 image (vllm-project/vllm#45381). Serves from the launch_b300-nv.sh MODEL_PATH split (unstaged model -> writable /data/models). Co-Authored-By: Claude Fable 5 --- .github/configs/nvidia-master.yaml | 37 ++++++ .../fixed_seq_len/minimaxm3_fp8_b300.sh | 109 ++++++++++++++++++ 2 files changed, 146 insertions(+) create mode 100755 benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index d22ae3229..ef5ed691e 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -11678,3 +11678,40 @@ minimaxm2.5-fp8-gb300-dynamo-vllm: tp: 4 ep: 4 dp-attn: true + +# MiniMax-M3 day-zero (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3). +# 427B total / 26B active MoE with MSA sparse attention; MXFP8 checkpoint +# (MiniMaxAI/MiniMax-M3-MXFP8, ~444 GB) quantized by NVIDIA — native MX tensor +# cores on Blackwell. M3 support has not shipped in a stable vLLM release; +# the dedicated vllm/vllm-openai:minimax-m3 image is built from the m3_release +# branch (vllm-project/vllm#45381). --block-size 128 is mandatory (MSA +# sparse/index cache alignment). Weights are NOT SRE-staged; b300 falls back +# to writable /data/models (see launch_b300-nv.sh MODEL_PATH split). +minimaxm3-fp8-b300-vllm: + 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: 4, conc-end: 64 } + - { tp: 8, ep: 8, conc-start: 128, conc-end: 512 } + - { tp: 4, conc-start: 4, conc-end: 64 } + - { tp: 4, ep: 4, conc-start: 128, conc-end: 512 } + # tp2 fits MXFP8 weights (~222 GB/GPU of 288) but KV headroom is thin; + # 1k1k only, drop if it OOMs at the high end. + - { tp: 2, ep: 2, conc-start: 16, conc-end: 128 } + - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 1024 } + - isl: 8192 + osl: 1024 + search-space: + - { tp: 8, conc-start: 4, conc-end: 64 } + - { tp: 8, ep: 8, conc-start: 128, conc-end: 512 } + - { tp: 4, conc-start: 4, conc-end: 128 } + - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 512 } diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh new file mode 100755 index 000000000..9d19d3724 --- /dev/null +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh @@ -0,0 +1,109 @@ +#!/usr/bin/env bash + +# MiniMax-M3 MXFP8 B300 single-node vLLM recipe +# (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3). Same shape as the B200 +# script, but follows the b300 launcher's MODEL/MODEL_PATH split: +# launch_b300-nv.sh keeps MODEL as the HF id and points MODEL_PATH at +# /data/models/ (writable NFS) for models not in the SRE-staged +# /scratch/models list — MiniMax-M3 is not staged. --block-size 128 is +# mandatory (MSA sparse/index cache); the benchmark is text-only, so +# --language-model-only frees the vision encoder's VRAM. + +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="$MODEL" +fi + +if [[ -n "$SLURM_JOB_ID" ]]; then + echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME" +fi + +nvidia-smi + +SERVER_LOG=/workspace/server.log + +# 444 GB of MXFP8 weights off shared FS; engine startup can exceed the +# default 600s readiness window. +export VLLM_ENGINE_READY_TIMEOUT_S=3600 + +if [ "${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: capture +# up to the next power of two >= CONC, capped at vLLM's 2048 ceiling. +CAPTURE_SIZE=4 +while (( CAPTURE_SIZE < CONC )); 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_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 \ +--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 "$SERVER_PID" + +run_benchmark_serving \ + --model "$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 "$RESULT_FILENAME" \ + --result-dir /workspace/ \ + --trust-remote-code + +# After throughput, run evaluation only if RUN_EVAL is true +if [ "${RUN_EVAL}" = "true" ]; then + run_eval --framework lm-eval --port "$PORT" + append_lm_eval_summary +fi + +# Stop GPU monitoring +stop_gpu_monitor +set +x From 512a37601a28920f30010ebf820746d2add2cb26 Mon Sep 17 00:00:00 2001 From: Cam Quilici Date: Fri, 12 Jun 2026 14:54:45 -0500 Subject: [PATCH 2/4] minimaxm3-fp8-b300-vllm: add perf-changelog entry Co-Authored-By: Claude Fable 5 --- perf-changelog.yaml | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index fd6b08dee..d29c9a5d3 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3636,3 +3636,13 @@ - "Add MiniMax-M2.5 FP4 (NVFP4) B300 TensorRT-LLM benchmark (model: nvidia/MiniMax-M2.5-NVFP4)" - "Image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1712 + +- config-keys: + - minimaxm3-fp8-b300-vllm + description: + - "Initial submission: MiniMax-M3 MXFP8 day-zero single-node vLLM benchmark on B300 (model: MiniMaxAI/MiniMax-M3-MXFP8, 427B total / 26B active MoE with MSA sparse attention)" + - "Image: vllm/vllm-openai:minimax-m3 (already the cu130 build; M3 support is unmerged upstream — vllm-project/vllm#45381)" + - "--block-size 128 is mandatory (MSA sparse/index cache alignment); --language-model-only skips the vision encoder for text-only throughput; conc-scaled --max-cudagraph-capture-size" + - "Layouts: TP8 and TP4 (latency), TP4+EP4 / TP8+EP8 (TEP throughput), tp2-ep2, TP8+EP8 dp-attn (DEP) across 1k1k and 8k1k" + - "Serves from the launch_b300-nv.sh MODEL/MODEL_PATH split (model not in the SRE-staged /scratch/models list -> writable /data/models)" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1724 From a50e22f5e029a4c1ef6521290eba852086f8f535 Mon Sep 17 00:00:00 2001 From: Cam Quilici Date: Fri, 12 Jun 2026 15:32:05 -0500 Subject: [PATCH 3/4] minimaxm3-fp8-b300-vllm: extend TEP8 to low conc for latency frontier Address PR #1724 review: TP8+EP8 conc-start 128->4 (1k1k and 8k1k) to probe whether TEP8 extends the min-latency frontier below plain TP8; TP4+EP4 conc-start 128->64 (1k1k) to fill the mid-curve. Co-Authored-By: Claude Fable 5 --- .github/configs/nvidia-master.yaml | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index ef5ed691e..d4313b76e 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -11701,9 +11701,9 @@ minimaxm3-fp8-b300-vllm: osl: 1024 search-space: - { tp: 8, conc-start: 4, conc-end: 64 } - - { tp: 8, ep: 8, conc-start: 128, conc-end: 512 } + - { tp: 8, ep: 8, conc-start: 4, conc-end: 512 } - { tp: 4, conc-start: 4, conc-end: 64 } - - { tp: 4, ep: 4, conc-start: 128, conc-end: 512 } + - { tp: 4, ep: 4, conc-start: 64, conc-end: 512 } # tp2 fits MXFP8 weights (~222 GB/GPU of 288) but KV headroom is thin; # 1k1k only, drop if it OOMs at the high end. - { tp: 2, ep: 2, conc-start: 16, conc-end: 128 } @@ -11712,6 +11712,6 @@ minimaxm3-fp8-b300-vllm: osl: 1024 search-space: - { tp: 8, conc-start: 4, conc-end: 64 } - - { tp: 8, ep: 8, conc-start: 128, conc-end: 512 } + - { tp: 8, ep: 8, conc-start: 4, conc-end: 512 } - { tp: 4, conc-start: 4, conc-end: 128 } - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 512 } From 1ccf751c625d57865dfab576867e60e861de13f6 Mon Sep 17 00:00:00 2001 From: Cam Quilici Date: Fri, 12 Jun 2026 16:14:04 -0500 Subject: [PATCH 4/4] minimaxm3-fp8-b300-vllm: add conc 1 and 2 to latency layouts Lower conc-start 4->1 on the latency-probing layouts (tp8, tp8+ep8, tp4) for both 1k1k and 8k1k to capture single/dual-request min-latency points. Co-Authored-By: Claude Fable 5 --- .github/configs/nvidia-master.yaml | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index d4313b76e..187824347 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -11700,9 +11700,9 @@ minimaxm3-fp8-b300-vllm: - isl: 1024 osl: 1024 search-space: - - { tp: 8, conc-start: 4, conc-end: 64 } - - { tp: 8, ep: 8, conc-start: 4, conc-end: 512 } - - { tp: 4, conc-start: 4, conc-end: 64 } + - { tp: 8, conc-start: 1, conc-end: 64 } + - { tp: 8, ep: 8, conc-start: 1, conc-end: 512 } + - { tp: 4, conc-start: 1, conc-end: 64 } - { tp: 4, ep: 4, conc-start: 64, conc-end: 512 } # tp2 fits MXFP8 weights (~222 GB/GPU of 288) but KV headroom is thin; # 1k1k only, drop if it OOMs at the high end. @@ -11711,7 +11711,7 @@ minimaxm3-fp8-b300-vllm: - isl: 8192 osl: 1024 search-space: - - { tp: 8, conc-start: 4, conc-end: 64 } - - { tp: 8, ep: 8, conc-start: 4, conc-end: 512 } - - { tp: 4, conc-start: 4, conc-end: 128 } + - { tp: 8, conc-start: 1, conc-end: 64 } + - { tp: 8, ep: 8, conc-start: 1, conc-end: 512 } + - { tp: 4, conc-start: 1, conc-end: 128 } - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 512 }