[AMD] Update MiniMax-M3-MXFP4 MI355X vLLM disagg perf and config

.github/configs/amd-master.yaml

@@ -2602,9 +2602,9 @@ minimaxm3-fp4-mi355x-vllm-disagg:
     - isl: 8192
       osl: 1024
       search-space:
-      # 1P TP4 + 1D TP4 (2 nodes total), conc sweep 1..512 (single job, looped)
+      # 1P TP4 + 1D TP4 (2 nodes total), conc sweep 1..256 (single job, looped)
       - spec-decoding: "none"
-        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128, 256, 512 ]
+        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128, 256 ]
         prefill:
           num-worker: 1
           tp: 4

benchmarks/multi_node/amd_utils/models_vllm.yaml

@@ -42,6 +42,7 @@ MiniMax-M3-MXFP4:
   prefill_flags: "--tensor-parallel-size 8 --max-num-batched-tokens 32768 --max-num-seqs 512 --block-size 128 --language-model-only --attention-backend TRITON_ATTN --moe-backend aiter --no-enable-prefix-caching --gpu-memory-utilization 0.90 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice"
   decode_flags: "--tensor-parallel-size 8 --max-num-batched-tokens 32768 --max-num-seqs 512 --block-size 128 --language-model-only --attention-backend TRITON_ATTN --moe-backend aiter --no-enable-prefix-caching --gpu-memory-utilization 0.90 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice"
   env: "VLLM_USE_V1=1 VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_MOE=1 VLLM_USE_BREAKABLE_CUDAGRAPH=0 VLLM_ENGINE_READY_TIMEOUT_S=3600"
+  prefill_env: "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB=2048"
   hf_dir: "models--amd--MiniMax-M3-MXFP4"
 
 gpt-oss-120b:

benchmarks/multi_node/amd_utils/server_vllm.sh

@@ -145,10 +145,12 @@ pf = bash_escape(m.get('prefill_flags', '--tensor-parallel-size 8'))
 df = bash_escape(m.get('decode_flags', '--tensor-parallel-size 8'))
 ev = bash_escape(m.get('env', ''))
 dev = bash_escape(m.get('decode_env', ''))
+pev = bash_escape(m.get('prefill_env', ''))
 print(f'PREFILL_SERVER_CONFIG=\"{pf}\"')
 print(f'DECODE_SERVER_CONFIG=\"{df}\"')
 print(f'MODEL_ENVS=\"{ev}\"')
 print(f'DECODE_MODEL_ENVS=\"{dev}\"')
+print(f'PREFILL_MODEL_ENVS=\"{pev}\"')
 ")"
 
 echo "Loaded model configuration for: $MODEL_NAME"
@@ -251,6 +253,11 @@ if [ "$NODE_RANK" -eq 0 ]; then
 
     setup_vllm_env
 
+    for env_pair in ${PREFILL_MODEL_ENVS}; do
+        export "$env_pair"
+        echo "[PREFILL_ENV] $env_pair"
+    done
+
     # Router is started as an external container by job.slurm (VLLM_ROUTER_IMAGE)
     echo "Using external vllm-router container (started by job.slurm on this node)"
 
@@ -420,6 +427,11 @@ elif [ "$NODE_RANK" -gt 0 ] && [ "$NODE_RANK" -lt "$xP" ]; then
 
     setup_vllm_env
 
+    for env_pair in ${PREFILL_MODEL_ENVS}; do
+        export "$env_pair"
+        echo "[PREFILL_ENV] $env_pair"
+    done
+
     SERVED_MODEL="${MODEL_NAME}"
     PREFILL_CMD="vllm serve ${MODEL_PATH} \
         --served-model-name ${SERVED_MODEL} \

perf-changelog.yaml

@@ -4255,3 +4255,10 @@
     - "Reuse the pinned vllm/vllm-openai-rocm:nightly-3f5a1e1733200760169ff31ebe60a271072b199e image, text-only target path, TRITON_ATTN, automatic tool choice, MiniMax-M3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0, default KV-cache dtype, and automatic MoE backend selection."
     - "Pass --use-chat-template for MTP acceptance and mirror the existing MiniMax-M3 MXFP8 MI355X MTP TP/EP/DP-attention search space at 1k1k and 8k1k."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1939
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-vllm-disagg
+  description:
+    - "Enable prefill-only INT4 quick-reduce for MiniMax-M3 MXFP4 disagg: set VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 and VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB=2048 on the prefill workers via a new prefill_env channel (mirrors the existing decode_env path in server_vllm.sh)."
+    - "Cap the 1P1D TP4 concurrency sweep at 256 (was 512); 2P1D TP4 unchanged at 128/256/512."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1943