[AMD] Tune MiniMax-M3 MXFP8 MI300X vLLM: async scheduling + big-prefill, fix conc256 EP8→EP1

.github/configs/amd-master.yaml

@@ -2877,7 +2877,7 @@ minimaxm3-fp4-mi355x-atom-disagg:
 # checkpoint lacks calibrated ROCm FP8 attention scales. Use the TP8-only H100
 # search space: TP8 for latency and TP8+EP8 (TEP) at high concurrency.
 minimaxm3-fp8-mi300x-vllm:
-  image: vllm/vllm-openai-rocm:minimax-m3
+  image: vllm/vllm-openai-rocm:nightly-4559c43a9526597c00cbcc4f59979496500268d1
   model: MiniMaxAI/MiniMax-M3-MXFP8
   model-prefix: minimaxm3
   runner: mi300x
@@ -2889,8 +2889,8 @@ minimaxm3-fp8-mi300x-vllm:
     - isl: 1024
       osl: 1024
       search-space:
-      - { tp: 8, conc-start: 1, conc-end: 128 }
-      - { tp: 8, ep: 8, conc-start: 256, conc-end: 256 }
+      # TP8/EP1 full conc range; the prior EP8 conc256 row regressed throughput.
+      - { tp: 8, conc-start: 1, conc-end: 256 }
     - isl: 8192
       osl: 1024
       search-space:

benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi300x.sh

@@ -5,6 +5,9 @@
 # is mandatory for MSA sparse attention. Keep the default BF16 KV cache on
 # gfx942: the checkpoint has no calibrated q/prob scales for ROCm FP8
 # attention, and vLLM's fallback scale of 1.0 corrupts model accuracy.
+# --async-scheduling and --max-num-batched-tokens 16384 are accuracy-safe
+# scheduling levers: the larger prefill batch amortizes the per-step
+# BF16-emulated MoE weight read, lifting high-concurrency throughput.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
@@ -34,6 +37,13 @@ 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_MHA=0
+
+export TORCH_BLAS_PREFER_HIPBLASLT=1
+export NCCL_MIN_NCHANNELS="${NCCL_MIN_NCHANNELS:-112}"
+export GPU_MAX_HW_QUEUES="${GPU_MAX_HW_QUEUES:-2}"
+
 if [ "${EVAL_ONLY}" = "true" ]; then
     setup_eval_context
 fi
@@ -59,6 +69,8 @@ vllm serve "$MODEL" --port "$PORT" \
     --language-model-only \
     --max-model-len "$MAX_MODEL_LEN" \
     --attention-backend TRITON_ATTN \
+    --async-scheduling \
+    --max-num-batched-tokens 16384 \
     --tool-call-parser minimax_m3 \
     --reasoning-parser minimax_m3 \
     --enable-auto-tool-choice > "$SERVER_LOG" 2>&1 &

perf-changelog.yaml

@@ -4324,6 +4324,17 @@
     - "Pin vllm/vllm-openai-rocm:nightly-4559c43a9526597c00cbcc4f59979496500268d1 (from nightly-3f5a1e1733200760169ff31ebe60a271072b199e) for AITER MoE and shared-expert fusion support (vllm-project/vllm#46419, vllm-project/vllm#46545)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1954
 
+- config-keys:
+    - minimaxm3-fp8-mi300x-vllm
+  description:
+    - "Update the MiniMax-M3 MXFP8 MI300X vLLM image from vllm/vllm-openai-rocm:minimax-m3 to vllm/vllm-openai-rocm:nightly-4559c43a9526597c00cbcc4f59979496500268d1."
+    - "Enable AITER kernels for MiniMax-M3 MXFP8 on MI300X/gfx942 via the single master toggle VLLM_ROCM_USE_AITER=1: the stock image left it unset, so the hot decode GEMMs and fused MoE ran on the generic kernels. The per-component AITER flags (MoE, linear, RMSNorm, FP8 batched-GEMM) default to True and are gated behind the master flag, so they are left at their defaults. Keep attention on TRITON_ATTN (VLLM_ROCM_USE_AITER_MHA=0, which defaults to True) because the MXFP8 checkpoint lacks calibrated q/prob scales for ROCm FP8 attention."
+    - "Add AMD-recommended, numerically-inert MI300X runtime knobs: TORCH_BLAS_PREFER_HIPBLASLT=1, NCCL_MIN_NCHANNELS=112 (raises RCCL channels above the ~32-64 default for TP8), GPU_MAX_HW_QUEUES=2 (caps HIP streams below the default of 4)."
+    - "Stack two accuracy-safe scheduling levers: --async-scheduling (overlaps CPU input-prep with GPU decode) and --max-num-batched-tokens 16384 (amortizes the per-step BF16-emulated MoE weight read of ~95 GB/rank over more prompt tokens, halving prefill weight-reads vs the 8192 default). Both are token-for-token identical (scheduling only); GSM8K exact-match holds at 0.959."
+    - "Switch the 1k1k conc256 search-space row from TP8/EP8 to TP8/EP1: the EP8 topology regressed high-concurrency throughput (434 vs 905 tok/s/gpu @ conc256, EP8 vs EP1) and EP1 matches the topology the prior AITER uplift was measured against."
+    - "Measured uplift on 8xMI300X, 1k1k random sweep (total tok/s/gpu): conc256 782.7->905 (EP8->EP1 + scheduling levers), conc128 598.9->628 (+4.9%), conc64 365.1->429 (+17.5%), conc32 295.6->327.4 (+10.8%), conc16 203.1->216.5 (+6.6%), conc8 127.6->136.6 (+7.1%), conc4 80.1->84.6 (+5.6%); conc1-2 unchanged (latency-bound). GSM8K exact-match holds at 0.959."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1951
+
 - config-keys:
     - minimaxm3-fp8-mi355x-vllm
     - minimaxm3-fp8-mi355x-vllm-mtp