sqrt-alibi flash attention算子是否有 triton 实现？

[zsc]

国产芯片适配需要

[yangxueruivs]

triton_unified_attention.py
try this.

[winterslong]

triton_unified_attention.py try this.

我们最近在H20上测试了VLLM 0.15.0（pr#32511）+Step-Audio-Chat模型，推理结果为空，能否看看有什么问题吗？

`

stepfun-ai/Step-Audio# vllm serve /stepfun-ai/checkpoints/Step-Audio-Chat --dtype auto -tp 4 --served-model-name step-audio-chat --trust-remote-code --attention-backend TRITON_ATTN

stepfun-ai/Step-Audio# python call_vllm_chat.py

completion:ChatCompletion(id='chatcmpl-a7d41d5306210af2', choices=[Choice(finish_reason='stop', index=0, logprobs=None, message=ChatCompletionMessage(content='', refusal=None, role='assistant', annotations=None, audio=None, function_call=None, tool_calls=[], reasoning=None, reasoning_content=None), stop_reason=3, token_ids=None)], created=1769999701, model='step-audio-chat', object='chat.completion', service_tier=None, system_fingerprint=None, usage=CompletionUsage(completion_tokens=1, prompt_tokens=32, total_tokens=33, completion_tokens_details=None, prompt_tokens_details=None), prompt_logprobs=None, prompt_token_ids=None, kv_transfer_params=None)
output:`