[AMD] MiniMax-M3 FP4 MI355X vLLM STP: INT4 all-reduce + fp8 KV, TP4 sweep / MiniMax-M3 FP4 MI355X vLLM STP：INT4 all-reduce + fp8 KV，TP4 扫描

benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_vllm.sh

@@ -4,6 +4,12 @@
 # https://huggingface.co/amd/MiniMax-M3-MXFP4#reproduction
 # Block size 128 is mandatory for MSA. This fixed-sequence benchmark uses the
 # text-only language-model path with AITER MoE (vllm-project/vllm#46419).
+#
+# High-concurrency tuning (no model-architecture overrides):
+#   * INT4 quantized all-reduce (env knobs below) -- reduces the all-reduce
+#     cost (the biggest decode kernel); measured ~-12% to -17% TPOT at conc
+#     64/128/256.
+#   * fp8 KV cache (--kv-cache-dtype fp8).
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
@@ -35,6 +41,13 @@ export VLLM_USE_BREAKABLE_CUDAGRAPH=0
 export VLLM_ROCM_USE_AITER=1
 export VLLM_ROCM_USE_AITER_MOE=1
 export VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1
+# INT4 quantized all-reduce for the (~1.5 MB) decode all-reduces, which are the
+# single biggest decode kernel at high concurrency. The MIN_SIZE_KB override is
+# required: vLLM's default INT4 quick-reduce size gate for (bf16, TP4) is 16 MB,
+# so it never fires for decode-sized tensors without it.
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+export VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16=0
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION_MIN_SIZE_KB=256
 
 if [ "${EVAL_ONLY}" = "true" ]; then
     setup_eval_context
@@ -63,6 +76,7 @@ vllm serve "$MODEL" --port "$PORT" \
     --max-model-len "$MAX_MODEL_LEN" \
     --attention-backend TRITON_ATTN \
     --moe-backend aiter \
+    --kv-cache-dtype fp8 \
     --tool-call-parser minimax_m3 \
     --enable-auto-tool-choice \
     --reasoning-parser minimax_m3 > "$SERVER_LOG" 2>&1 &

configs/amd-master.yaml

@@ -2472,7 +2472,7 @@ minimaxm3-fp4-mi355x-vllm-disagg:
 # language-model path and mirror the MXFP8 MI355X search space for a direct
 # precision comparison.
 minimaxm3-fp4-mi355x-vllm:
-  image: vllm/vllm-openai-rocm:nightly-4559c43a9526597c00cbcc4f59979496500268d1
+  image: vllm/vllm-openai-rocm:nightly-69715823df89b11ee684b84066390cbb9092d5c1
   model: amd/MiniMax-M3-MXFP4
   model-prefix: minimaxm3
   runner: mi355x
@@ -2484,19 +2484,11 @@ minimaxm3-fp4-mi355x-vllm:
     - 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 }
+      - { tp: 4, conc-start: 1, conc-end: 512 }
     - 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 }
+      - { tp: 4, conc-start: 1, conc-end: 512 }
 
 # EAGLE3 speculative-decoding variant of minimaxm3-fp4-mi355x-vllm. Pair the
 # amd/MiniMax-M3-MXFP4 target with Inferact/MiniMax-M3-EAGLE3 and three draft

perf-changelog.yaml

@@ -4574,3 +4574,11 @@
   description:
     - "Add high concurrency configs"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1994
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-vllm
+  description:
+    - "Close the high-concurrency gap vs the ATOM recipe on MiniMax-M3 MXFP4 MI355X single-node vLLM STP. Bump image to latest nightly vllm/vllm-openai-rocm:nightly-69715823df89b11ee684b84066390cbb9092d5c1."
+    - "Add INT4 quantized all-reduce (VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4, VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16=0, VLLM_ROCM_QUICK_REDUCE_QUANTIZATION_MIN_SIZE_KB=256) and fp8 KV cache (--kv-cache-dtype fp8). No --hf-overrides indexer-skipping knobs (disallowed: reduces model architecture FLOPs per PR_REVIEW_CHECKLIST.md)."
+    - "Simplify search space to a single TP4 conc 1-512 sweep for 1k1k and 8k1k (drop TP8, TP4/EP4, TP2/EP2, and dp-attn variants), matching minimaxm3-fp8-mi355x-vllm (#2003)."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2104