[Feature request] Support attention logits cap with tanh

[merrymercy]

The grok model uses tanh to cap the attention logits. Could you support this feature in flashinfer? If you need community help, any instructions on how to add this will be appreciated.

Grok (jax): https://github.com/xai-org/grok-1/blob/7050ed204b8206bb8645c7b7bbef7252f79561b0/model.py#L864-L865

SGLang implementation (triton): https://github.com/sgl-project/sglang/blob/2cea6146d8735780da602c0dfa0569b0fb5d47ba/python/sglang/srt/layers/extend_attention.py#L101-L102

[yzh119]

Sounds good, should be easy to support.

[yzh119]

Is there is formal name with this "Attention with Logits Cap" method?

[merrymercy]

there is no formal name. maybe just call it "logit cap"

[merrymercy]

@yzh119 Any progress on this issue? FYI, TensorRT-LLM recently added this feature https://github.com/NVIDIA/TensorRT-LLM/blob/db4edea1e1359bcfcac7bbb87c1b639b5611c721/tensorrt_llm/functional.py#L4519-L4521

[yzh119]

Done in #298