[rocm]use aiter triton kernel as triton mha fallback path

[zhuyuhua-v]

Purpose

use aiter triton kernel as triton mha fallback path instead of aiter fmha kernel.

Test Plan

server:

export VLLM_USE_V1=1
export SAFETENSORS_FAST_GPU=1
export VLLM_ROCM_USE_AITER=1
export VLLM_ROCM_USE_AITER_RMSNORM=1
export VLLM_ROCM_USE_AITER_MOE=1
export VLLM_USE_TRITON_FLASH_ATTN=1
export NCCL_DEBUG=WARN
export VLLM_RPC_TIMEOUT=1800000
export VLLM_ROCM_USE_AITER_MHA=0
export VLLM_ROCM_USE_TRITON_ROPE=1
export VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1

export VLLM_ROCM_USE_AITER_MLA=0 # triton path

model_path="path_to_model/deepseek-ai/DeepSeek-V3"
vllm serve $model_path \
    --tensor-parallel-size 8 \
    --max-num-batched-tokens 32768 \
    --trust-remote-code \
    --no-enable-prefix-caching \
    --disable-log-requests \
    --gpu_memory_utilization 0.9 \
    --port 6789 \
    --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
    --block-size 16 \
    --async-scheduling \
    --enforce-eager \

accuracy:

model="path_to_model/deepseek-ai/DeepSeek-V3"
lm_eval \
--model local-completions \
--tasks gsm8k \
--seed 123 \
--model_args model=${model},base_url=http://127.0.0.1:6789/v1/completions \
--batch_size 100 \

Test Result

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.9477	±	0.0061
		strict-match	5	exact_match	↑	0.9454	±	0.0063