[AMD] Add MiniMax-M3-MXFP4 MI355X single-node vLLM recipe

.github/configs/amd-master.yaml

@@ -2587,6 +2587,32 @@ 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 single-node vLLM recipe.
+# amd/MiniMax-M3-MXFP4 via the AITER MoE backend (block size 128 for MSA, no fp8
+# KV cache -- the checkpoint has no calibrated KV scales). Upstream recipe:
+# https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3 (mxfp4 variant).
+# TP8 covers the low-latency / low-concurrency regime; TP4 carries the full sweep.
+minimaxm3-fp4-mi355x-vllm:
+  image: rocm/vllm-dev:vllm-0.23.1-rocm723-mi35x-mori-0625
+  model: amd/MiniMax-M3-MXFP4
+  model-prefix: minimaxm3
+  runner: mi355x
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 1, conc-end: 32 }
+      - { tp: 4, conc-start: 1, conc-end: 256 }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 1, conc-end: 32 }
+      - { tp: 4, conc-start: 1, conc-end: 256 }
+
 # 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.

benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x.sh

@@ -0,0 +1,91 @@
+#!/usr/bin/env bash
+
+# MiniMax-M3 MXFP4 MI355X (gfx950) single-node vLLM recipe.
+# amd/MiniMax-M3-MXFP4 served through the AITER MoE backend. Requires block size
+# 128 (MSA sparse attention). KV cache is left at the default dtype: this
+# checkpoint ships no calibrated FP8 KV scales, so --kv-cache-dtype fp8 would
+# fall back to an uncalibrated scale of 1.0 (accuracy risk).
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    EP_SIZE \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+SERVER_LOG=/workspace/server.log
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+export VLLM_USE_BREAKABLE_CUDAGRAPH=0
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_USE_AITER_MOE=1
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+fi
+
+PARALLEL_ARGS=(--tensor-parallel-size "$TP")
+if [ "${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+=(--enable-expert-parallel)
+fi
+
+start_gpu_monitor
+
+set -x
+vllm serve "$MODEL" --port "$PORT" \
+    "${PARALLEL_ARGS[@]}" \
+    --block-size 128 \
+    --no-enable-prefix-caching \
+    --language-model-only \
+    --max-model-len "$MAX_MODEL_LEN" \
+    --moe-backend aiter \
+    --attention-backend TRITON_ATTN \
+    --tool-call-parser minimax_m3 \
+    --reasoning-parser minimax_m3 \
+    --enable-auto-tool-choice > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x

perf-changelog.yaml

@@ -4190,3 +4190,9 @@
     - "Update the DeepSeek-V4-Pro B200 disaggregated Dynamo-vLLM benchmark to the vllm/vllm-openai:v0.23.0 image"
     - "Lower max-num-batched-tokens to 16384 and gpu-memory-utilization to 0.9 on the high-throughput and max-throughput recipes to avoid OOM"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1899
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-vllm
+  description:
+    - "Initial submission: MiniMax-M3 MXFP4 single-node on MI355X with vLLM (AITER MoE, TRITON_ATTN MSA, block-size 128, no fp8 KV cache). Pure TP: TP8 low-concurrency (conc 1-32) + TP4 full sweep (conc 1-256), at 1k/1k and 8k/1k."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1936