[Do Not Merge] [WIP till AMD releases MXFP4 of MiniMax M2.5] Add MiniMax M2.1 MXFP4 benchmark for MI355x vLLM (TP=2,4)

.github/configs/amd-master.yaml

@@ -314,6 +314,31 @@ kimik2.5-fp4-mi355x-vllm:
     search-space:
     - { tp: 8, conc-start: 4, conc-end: 64 }
 
+minimaxm2.1-fp4-mi355x-vllm:
+  image: vllm/vllm-openai-rocm:v0.16.0
+  model: amd/MiniMax-M2.1-MXFP4
+  model-prefix: minimaxm2.1
+  runner: mi355x
+  precision: fp4
+  framework: vllm
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 2, conc-start: 4, conc-end: 64 }
+    - { tp: 4, conc-start: 4, conc-end: 64 }
+  - isl: 1024
+    osl: 8192
+    search-space:
+    - { tp: 2, conc-start: 4, conc-end: 64 }
+    - { tp: 4, conc-start: 4, conc-end: 64 }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 2, conc-start: 4, conc-end: 64 }
+    - { tp: 4, conc-start: 4, conc-end: 64 }
+
 minimaxm2.5-fp8-mi355x-vllm:
   image: vllm/vllm-openai-rocm:v0.15.1
   model: MiniMaxAI/MiniMax-M2.5

benchmarks/single_node/minimaxm2.1_fp4_mi355x.sh

@@ -0,0 +1,71 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+hf download "$MODEL"
+
+# Set HIP_VISIBLE_DEVICES to match ROCR_VISIBLE_DEVICES for Ray compatibility in vLLM 0.14+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+export VLLM_ROCM_USE_AITER=1
+if [ "$TP" -ge 4 ]; then
+    # AITER CK fused MoE kernels lack compiled tiles for N=intermediate_size/TP
+    # when TP>=4 (TP=4, N=384). Disable AITER MoE to fall back to triton, but keep
+    # AITER attention. See: https://github.com/vllm-project/vllm/issues/35637
+    export VLLM_ROCM_USE_AITER_MOE=0
+    export VLLM_ATTENTION_BACKEND="ROCM_AITER_UNIFIED_ATTN"
+    pip install amd-quark 2>/dev/null || true
+fi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+set -x
+vllm serve $MODEL --port $PORT \
+--tensor-parallel-size=$TP \
+--gpu-memory-utilization 0.95 \
+--max-model-len $MAX_MODEL_LEN \
+--block-size=32 \
+--disable-log-requests \
+--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" --concurrent-requests $CONC
+    append_lm_eval_summary
+fi
+set +x

perf-changelog.yaml

@@ -755,6 +755,17 @@
     - "Key changes: AITER v0.1.10.post3 with FP8 Prefill/Decode/KV Cache, FP8 prefill attention kernel, MORI EP two-batch overlapping, OOM fix for DeepSeek weight loading"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/816
 
+- config-keys:
+    - minimaxm2.1-fp4-mi355x-vllm
+  description:
+    - "Add MiniMax M2.1 MXFP4 vLLM benchmark for MI355X"
+    - "Model: amd/MiniMax-M2.1-MXFP4 with --trust-remote-code and --block-size=32"
+    - "Image: vllm/vllm-openai-rocm:v0.16.0"
+    - "Environment: VLLM_ROCM_USE_AITER=1"
+    - "TP=2 only (TP=4 disabled due to vLLM bug https://github.com/vllm-project/vllm/issues/35637)"
+    - "Concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/827
+
 - config-keys:
     - minimaxm2.5-fp8-h200-vllm
   description: