[Performance][ROCm] Integrate aiter indexer scoring and top-k kernels into MiniMax-M3 sparse attention path

[ykamiset]

Integrate aiter indexer scoring and top-k kernels (AITER PR: ROCm/aiter#4787) into MiniMax-M3 sparse attention path.

Please note that this PR is dependent on these PRs:

• [ROCm][PERF] Enable AITER PA gluon decode for MiniMax-M3 MTP and dense layers #52849 current PR rebased on this PR- since it is dependent on it.
• [HIP] [JIT] [gfx950] Optimize Minimax M3 scoring & top-k kernels ROCm/aiter#4787

Test Plan

Server cmd to run

export VLLM_ENGINE_READY_TIMEOUT_S=3600
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=3600
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
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
export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1

vllm serve /model/MiniMax-M3-MXFP4 \
  --served-model-name MiniMaxAI/MiniMax-M3 \
  --host 0.0.0.0 \
  --port 8190 \
  --tensor-parallel-size 4 \
  --trust-remote-code \
  --block-size 128 \
  --enable-chunked-prefill \
  --language-model-only \
  --no-enable-prefix-caching \
  --max-num-batched-tokens 65536 \
  --default-chat-template-kwargs '{"thinking_mode":"enabled"}' \
  --max-num-seqs 128 \
  --kv-cache-dtype fp8 \
  --attention-backend ROCM_AITER_UNIFIED_ATTN \
  --attention-config '{"indexer_kv_dtype": "fp8"}' \
  --moe-backend aiter \
  --reasoning-parser minimax_m3 

Test Result

Conc	Baseline TPUT (tok/s) (Triton BF16)	AITER FP8 indexer TPUT (tok/s)	TPUT % uplift	Baseline P90 TPOT (Triton BF16)	AITER FP8 indexer P90 TPOT	TPOT % uplift	Baseline P90 ITL (Triton BF16)	AITER FP8 indexer P90 ITL	ITL % uplift
4	22955.29	26012.77	13.32%	20.18	17.01	15.71%	10.64	9.10	14.47%
32	34596.12	41526.05	20.03%	130.17	107.49	17.42%	23.72	17.05	28.12%

Accuracy score:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.9591	±	0.0055
		strict-match	5	exact_match	↑	0.9598	±	0.0054