From 654878b16097b6de607710cb4c59ccb408dc91d2 Mon Sep 17 00:00:00 2001 From: shekhar Date: Wed, 17 Jun 2026 18:36:53 +0000 Subject: [PATCH 1/6] minimaxm3-fp4-mi355x-atom: day-zero MiniMax-M3 MXFP4 MI355X atom recipe MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Smoke-tested on MI355X (mia1-p01-g07): TP4 conc-1 1k1k served and benched clean (mean TPOT 6.8ms). KV cache left at default dtype — amd/MiniMax-M3-MXFP4 has no calibrated FP8 KV scales, so --kv_cache_dtype fp8 asserts in the MSA fused_qknorm kernel. --- .github/configs/amd-master.yaml | 22 +++++ .../minimaxm3_fp4_mi355x_atom.sh | 88 +++++++++++++++++++ perf-changelog.yaml | 8 +- 3 files changed, 117 insertions(+), 1 deletion(-) create mode 100644 benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml index 606e3c2af..83c330ab7 100644 --- a/.github/configs/amd-master.yaml +++ b/.github/configs/amd-master.yaml @@ -2851,6 +2851,28 @@ minimaxm3-fp8-mi355x-vllm-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 MXFP4 MI355X atom recipe: +# https://github.com/ROCm/ATOM/blob/5d42d49f9e4292e5b61475917e92e7ec1b1dacb7/recipes/MiniMax-M3.md +# block size 128 is mandatory for MSA. TP4 on a single gfx950 node, per the recipe. +minimaxm3-fp4-mi355x-atom: + image: rocm/atom-dev:M3 + model: amd/MiniMax-M3-MXFP4 + model-prefix: minimaxm3 + runner: mi355x + precision: fp4 + framework: atom + multinode: false + scenarios: + fixed-seq-len: + - isl: 1024 + osl: 1024 + search-space: + - { tp: 4, conc-start: 1, conc-end: 128 } + - isl: 8192 + osl: 1024 + search-space: + - { tp: 4, conc-start: 1, conc-end: 128 } + # MiniMax-M3 MXFP8 MI300X day-zero recipe. Reuse the dedicated ROCm image and # MI355X serving shape, but retain the default BF16 KV cache because this # checkpoint lacks calibrated ROCm FP8 attention scales. Use the TP8-only H100 diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh new file mode 100644 index 000000000..5e0c789fe --- /dev/null +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh @@ -0,0 +1,88 @@ +#!/usr/bin/env bash + +source "$(dirname "$0")/../../benchmark_lib.sh" + +check_env_vars \ + MODEL \ + TP \ + CONC \ + ISL \ + OSL \ + RANDOM_RANGE_RATIO \ + RESULT_FILENAME \ + EP_SIZE \ + DP_ATTENTION + +if [[ -n "$SLURM_JOB_ID" ]]; then + echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME" +fi + +echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION" + +SERVER_LOG=/workspace/server.log + +export OMP_NUM_THREADS=1 + +# Calculate max-model-len based on ISL and OSL +if [ "$ISL" = "1024" ] && [ "$OSL" = "1024" ]; then + CALCULATED_MAX_MODEL_LEN="" +else + CALCULATED_MAX_MODEL_LEN=" --max-model-len 10240 " +fi + +if [ "$EP_SIZE" -gt 1 ]; then + EP=" --enable-expert-parallel" +else + EP=" " +fi + +# Start GPU monitoring (power, temperature, clocks every second) +start_gpu_monitor +MEM_FRAC_STATIC=0.8 + +set -x + +# Flags follow the ATOM MiniMax-M3 MXFP4 recipe (FP4 on 4xMI355 section): +# https://github.com/ROCm/ATOM/blob/5d42d49f9e4292e5b61475917e92e7ec1b1dacb7/recipes/MiniMax-M3.md +# --block-size 128 is mandatory for MiniMax MSA. KV cache is left at the default +# dtype: amd/MiniMax-M3-MXFP4 ships no calibrated FP8 KV scales, so +# --kv_cache_dtype fp8 trips an assertion (k_scale is None) in the MSA +# fused_qknorm kernel during init. +python3 -m atom.entrypoints.openai_server \ + --model $MODEL \ + --server-port $PORT \ + -tp $TP \ + $CALCULATED_MAX_MODEL_LEN $EP \ + --block-size 128 \ + --gpu-memory-utilization $MEM_FRAC_STATIC \ + --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 "$SERVER_PID" + +export PYTHONDONTWRITEBYTECODE=1 +run_benchmark_serving \ + --model "$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 "$PORT" + append_lm_eval_summary +fi + +# Stop GPU monitoring +stop_gpu_monitor +set +x diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 1f88b47aa..7b011574d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3918,4 +3918,10 @@ - "This issue is now fixed in the latest TRTLLM release." - "Also update all configs for DSR1 TRTLLM FP8 to reflect latest released image usage" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1767 - + +- config-keys: + - minimaxm3-fp4-mi355x-atom + description: + - "Add day-zero MiniMax-M3 MXFP4 (amd/MiniMax-M3-MXFP4) single-node atom benchmark on MI355X, following the ROCm/ATOM MiniMax-M3 recipe (TP4, block size 128 for MSA, FP8 KV cache)." + pr-link: TBD + From 2d7158a99bf7947d13f95331f1ad896f1f9663c0 Mon Sep 17 00:00:00 2001 From: shekhar Date: Wed, 17 Jun 2026 19:19:03 +0000 Subject: [PATCH 2/6] minimaxm3-fp4-mi355x-atom: route amd/MiniMax-M3* weights to NFS cache --- runners/launch_mi355x-amds.sh | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/runners/launch_mi355x-amds.sh b/runners/launch_mi355x-amds.sh index 1f367f5c0..acfd4912a 100644 --- a/runners/launch_mi355x-amds.sh +++ b/runners/launch_mi355x-amds.sh @@ -242,8 +242,9 @@ else fi # MiniMax-M3 weights are not staged on the node-local /var/lib NVMe cache; - # they are pre-downloaded once to the NFS share instead. - if [[ "$MODEL" == MiniMaxAI/MiniMax-M3* ]]; then + # they are pre-downloaded once to the NFS share instead. Covers both the + # MiniMaxAI MXFP8 checkpoint and the amd MXFP4 atom checkpoint. + if [[ "$MODEL" == MiniMaxAI/MiniMax-M3* || "$MODEL" == amd/MiniMax-M3* ]]; then export HF_HUB_CACHE_MOUNT="/it-share/hf-hub-cache/" fi From 9a2b0f4e930e98569da8ebbb1d6f2f0db9d14df8 Mon Sep 17 00:00:00 2001 From: shekhar Date: Wed, 17 Jun 2026 19:23:09 +0000 Subject: [PATCH 3/6] minimaxm3-fp4-mi355x-atom: fill perf-changelog pr-link --- perf-changelog.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 7b011574d..8a094ec8d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3922,6 +3922,6 @@ - config-keys: - minimaxm3-fp4-mi355x-atom description: - - "Add day-zero MiniMax-M3 MXFP4 (amd/MiniMax-M3-MXFP4) single-node atom benchmark on MI355X, following the ROCm/ATOM MiniMax-M3 recipe (TP4, block size 128 for MSA, FP8 KV cache)." - pr-link: TBD + - "Add day-zero MiniMax-M3 MXFP4 (amd/MiniMax-M3-MXFP4) single-node atom benchmark on MI355X, following the ROCm/ATOM MiniMax-M3 recipe (TP4, block size 128 for MSA, default KV cache dtype)." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1812 From a68c303fb28b3945567485783af6d0bf98e97463 Mon Sep 17 00:00:00 2001 From: shekhar Date: Wed, 17 Jun 2026 19:32:03 +0000 Subject: [PATCH 4/6] minimaxm3-fp4-mi355x-atom: use matrix MAX_MODEL_LEN (isl+osl+256) --- .../fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-) diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh index 5e0c789fe..80f230f28 100644 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh @@ -11,7 +11,8 @@ check_env_vars \ RANDOM_RANGE_RATIO \ RESULT_FILENAME \ EP_SIZE \ - DP_ATTENTION + DP_ATTENTION \ + MAX_MODEL_LEN if [[ -n "$SLURM_JOB_ID" ]]; then echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME" @@ -23,11 +24,11 @@ SERVER_LOG=/workspace/server.log export OMP_NUM_THREADS=1 -# Calculate max-model-len based on ISL and OSL -if [ "$ISL" = "1024" ] && [ "$OSL" = "1024" ]; then - CALCULATED_MAX_MODEL_LEN="" -else - CALCULATED_MAX_MODEL_LEN=" --max-model-len 10240 " +# Use the matrix-supplied MAX_MODEL_LEN (isl + osl + 256). Eval-only jobs need a +# larger window for the eval prompts, so override it from the eval context. +if [ "${EVAL_ONLY}" = "true" ]; then + setup_eval_context + MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN" fi if [ "$EP_SIZE" -gt 1 ]; then @@ -52,7 +53,7 @@ python3 -m atom.entrypoints.openai_server \ --model $MODEL \ --server-port $PORT \ -tp $TP \ - $CALCULATED_MAX_MODEL_LEN $EP \ + --max-model-len $MAX_MODEL_LEN $EP \ --block-size 128 \ --gpu-memory-utilization $MEM_FRAC_STATIC \ --trust-remote-code \ From 771ed5917ca75d7249fe55c9829ed0bf8cd956ea Mon Sep 17 00:00:00 2001 From: andyluo7 Date: Wed, 17 Jun 2026 13:31:37 -0700 Subject: [PATCH 5/6] trigger full-sweep validation From 61a6a94ca417c7d9edc05853b379cfa6e94ec90c Mon Sep 17 00:00:00 2001 From: andyluo7 Date: Wed, 17 Jun 2026 13:34:29 -0700 Subject: [PATCH 6/6] perf-changelog: point minimaxm3-fp4-mi355x-atom pr-link to upstream PR #1813 --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 8a094ec8d..52efcee9a 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3923,5 +3923,5 @@ - minimaxm3-fp4-mi355x-atom description: - "Add day-zero MiniMax-M3 MXFP4 (amd/MiniMax-M3-MXFP4) single-node atom benchmark on MI355X, following the ROCm/ATOM MiniMax-M3 recipe (TP4, block size 128 for MSA, default KV cache dtype)." - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1812 + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1813