Add k_scale and v_scale to persistent attention

[Edenzzzz]

📌 Description

For consistency with the general call in SGLang
cc @yzh119

🔍 Related Issues

🚀 Pull Request Checklist

Thank you for contributing to FlashInfer! Before we review your pull request, please make sure the following items are complete.

✅ Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit (or used your preferred method).
• I have installed the hooks with pre-commit install.
• I have run the hooks manually with pre-commit run --all-files and fixed any reported issues.

If you are unsure about how to set up pre-commit, see the pre-commit documentation.

🧪 Tests

• Tests have been added or updated as needed.
• All tests are passing (unittest, etc.).

Reviewer Notes

[gemini-code-assist]

Summary of Changes

Hello @Edenzzzz, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request enhances the persistent attention mechanism by introducing optional k_scale and v_scale parameters. These parameters allow for dynamic scaling of key and value tensors, respectively, providing greater flexibility and ensuring consistency with external frameworks like SGLang.

Highlights

• API Extension for Persistent Attention: Added k_scale and v_scale as optional torch.Tensor arguments to the run method of the persistent attention function in flashinfer/attention.py.
• Key Scaling Implementation: Implemented the application of k_scale to the _sm_scale (softmax scale factor) if provided, affecting the scaling of the attention scores before the main kernel execution.
• Value Scaling Implementation: Implemented the application of v_scale to the final output tensor (out) of the attention operation. This includes a conditional check for low-precision output types (itemsize == 1) to ensure correct scaling by temporarily casting to float.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments or fill out our survey to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request adds k_scale and v_scale parameters to the BatchAttention.run method for scaling the key and value tensors, respectively. My review identified two issues in the Python implementation: a stateful bug where self._sm_scale is modified in-place, which could lead to incorrect results on subsequent runs, and a critical bug where an externally provided output tensor out is not updated correctly. I've provided suggestions to fix both issues.

[gemini-code-assist]

When the out tensor is provided by the caller, reassigning it with out = ... breaks the contract, as the original tensor passed by the user will not be updated. The update must be performed in-place to ensure the caller's tensor is modified as expected.

                out.copy_((out.to(float) * v_scale).to(out.dtype))

[gemini-code-assist]

Modifying the instance attribute self._sm_scale in-place can lead to incorrect behavior if run() is called multiple times, as the scaling factor will accumulate with each call. The effective scale should be calculated in a local variable within the run method.

        sm_scale = self._sm_scale
        if self._sm_scale is None:
            sm_scale = 1.0 / math.sqrt(head_dim_qk)
        if k_scale is not None:
            sm_scale *= k_scale

        # profiler_buffer is optional
        profiler_args = (profiler_buffer,) if self._use_profiler else ()

        self.module.run(
            self.float_workspace_buffer,
            self.int_workspace_buffer,
            self._plan_info,
            q,
            k_cache,
            v_cache,
            self._kv_indices,
            out,
            lse,
            self._mask_mode,
            TensorLayout[self._kv_layout].value,
            self._num_qo_heads,
            self._num_kv_heads,
            self._page_size,
            sm_scale,
            *profiler_args,
        )

[yyihuang]

I think it's better to keep only one pre-fused param as attn softmax scale, with some examples here:
https://github.com/flashinfer-ai/flashinfer/blob/main/flashinfer/decode.py#L1799-L1800
https://github.com/flashinfer-ai/flashinfer/blob/main/flashinfer/decode.py#L2052-L2053

[Edenzzzz]

However, the interface has been like this in SGLang

[yyihuang]

It's better to keep the fusion overhead at inference engine side, like sgl-project/sglang#7938

https://github.com/farazkh80/sglang/blob/fkhoubsirat-trtllm_gen_mla_sglang/python/sglang/srt/layers/attention/trtllm_mla_backend.py#L352

[yzh119]

Can you also implement fusing the multiply inside the kernel in the persistent kernel template?

Also, it will be good to specify an out_dtype, for quantized input such as fp8, we usually use bf16 as output data type. You can check

flashinfer/flashinfer/gemm.py

Lines 1198 to 1203 in 435f341

	if out is None:
	out = torch.empty(
	(a.shape[0], b.shape[1]),
	device=a.device,
	dtype=out_dtype,
	)

for reference.

[Edenzzzz]

In that case we need to pass in out_dtype during plan? Which can be troublesome