[AMD] feat: MiniMax M3 Day 0 support MI355X

.github/configs/amd-master.yaml

@@ -2780,3 +2780,33 @@ dsv4-fp4-mi355x-atom-disagg:
           dp-attn: false
           additional-settings:
           - "DECODE_NODES=1"
+
+# MiniMax-M3 MXFP8 MI355X recipe:
+# https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5
+# MXFP8 runs from TP=4 on gfx950; block size 128 is mandatory for MSA.
+minimaxm3-fp8-mi355x-vllm:
+  image: vllm/vllm-openai-rocm:minimax-m3
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: mi355x
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { 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 }
+      - { 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: 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 }

benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi355x.sh

@@ -0,0 +1,86 @@
+#!/usr/bin/env bash
+
+# MiniMax-M3 MXFP8 MI355X (gfx950) single-node vLLM recipe.
+# https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5
+# The recipe recommends MXFP8 from TP=4 on gfx950 and requires block size 128.
+
+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="$ROCR_VISIBLE_DEVICES"
+fi
+
+SERVER_LOG=/workspace/server.log
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+
+if [ "${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" \
+    "${PARALLEL_ARGS[@]}" \
+    --block-size 128 \
+    --language-model-only \
+    --max-model-len "$MAX_MODEL_LEN" \
+    --kv-cache-dtype fp8 \
+    --attention-backend TRITON_ATTN \
+    --enforce-eager \
+    --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

@@ -3706,3 +3706,14 @@
     - "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
+
+- config-keys:
+    - minimaxm3-fp8-mi355x-vllm
+  description:
+    - "Initial submission: MiniMax-M3 MXFP8 day-zero single-node vLLM benchmark on MI355X / gfx950 (model: MiniMaxAI/MiniMax-M3-MXFP8, 427B total / 26B active MoE with MSA sparse attention)"
+    - "Image: vllm/vllm-openai-rocm:minimax-m3"
+    - "Matches vllm-project/recipes@2a3728ed MI355X guidance with --block-size 128, --attention-backend TRITON_ATTN, --kv-cache-dtype fp8, MiniMax-M3 tool/reasoning parsers, and automatic tool choice"
+    - "MI355X deviations required for fixed-sequence text benchmarking: --language-model-only, scenario-specific --max-model-len, and --enforce-eager to bypass the M3 decode CUDA-graph assertion"
+    - "B300-parity layouts and concurrency ranges: TP8, TP8+EP8, TP4, TP4+EP4, TP2+EP2, and TP8+EP8 dp-attn (DEP) across 1k1k and 8k1k"
+    - "launch_mi355x-amds.sh routes M3 weights to NFS /it-share/hf-hub-cache instead of node-local /var/lib NVMe"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1725

runners/launch_mi355x-amds.sh

@@ -241,6 +241,12 @@ else
         export HF_HUB_CACHE_MOUNT="/it-share/hf-hub-cache/"
     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 [[ "$MODEL" == MiniMaxAI/MiniMax-M3* ]]; then
+        export HF_HUB_CACHE_MOUNT="/it-share/hf-hub-cache/"
+    fi
+
     SCRIPT_BASE="${EXP_NAME%%_*}_${PRECISION}_mi355x"
     SCRIPT_FW="benchmarks/single_node/${SCENARIO_SUBDIR:-fixed_seq_len/}${SCRIPT_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
     SCRIPT_FALLBACK="benchmarks/single_node/${SCENARIO_SUBDIR:-fixed_seq_len/}${SCRIPT_BASE}${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh"