Add flash attention 3 support for LightLLMKernelTextEncoder

[sitatec]

No description provided.

[gemini-code-assist]

Summary of Changes

Hello @sitatec, 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 LightLLMKernelTextEncoder by integrating support for Flash Attention 3, offering users a newer, potentially more optimized attention mechanism. Concurrently, it refines the fallback attention strategy by adopting SDPA, which promises better performance and seamless integration with torch.compile for scenarios where Flash Attention is not utilized. These changes aim to provide more flexible and efficient attention handling within the text encoder.

Highlights

• Flash Attention 3 Support: Introduced explicit support for 'flash_attention_3' as an attention implementation option within the LightLLMKernelTextEncoder, allowing users to specify this version for potentially enhanced performance.
• Documentation Update: Updated both English and Chinese beginner guides to inform users about the new 'flash_attention_3' option for the use_flash_attention_kernel parameter.
• Default Attention Improvement: Switched the default attention implementation from 'eager' to 'sdpa' (Scaled Dot Product Attention) when use_flash_attention_kernel is not enabled. This change significantly improves performance and maintains compatibility with torch.compile.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

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 by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

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 pull request 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

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 support for FlashAttention 3 in the LightLLMKernelTextEncoder. The changes include updating the attention implementation selection logic in qwen25_text_encoder_kernel.py and documenting the new option in the beginner guides. Additionally, the fallback attention implementation has been switched from eager to the more performant sdpa.

My review focuses on improving the clarity of the new documentation and suggesting a minor readability improvement to the implementation logic. Overall, the changes are good and the performance improvement from using sdpa is a welcome addition.

[gemini-code-assist]

The description for use_flash_attention_kernel could be more explicit about the accepted values for clarity. It's good that you've documented the new option, but we can make it even clearer for users.

Suggested change

	* `use_flash_attention_kernel`: Enable FlashAttention acceleration for Attention layers. By default flash_attention_2 will be used, but you can also use "use_flash_attention_kernel": "flash_attention_3".
	* `use_flash_attention_kernel`: Enable FlashAttention acceleration for Attention layers. Set to `true` to use `flash_attention_2` (the default when enabled), or set to the string `"flash_attention_3"` to use FlashAttention 3.

[gemini-code-assist]

For clarity, the description for use_flash_attention_kernel could be more explicit about the accepted values. It's good that you've documented the new option, but we can make it even clearer for users. This comment is a translation of the one for the English documentation.

Suggested change

	* `use_flash_attention_kernel`: 启用 FlashAttention 加速 Attention 层。 默认情况下将使用 flash_attention_2，但你也可以使用 “use_flash_attention_kernel”: “flash_attention_3”。
	* `use_flash_attention_kernel`: 启用 FlashAttention 加速 Attention 层。设置为 `true` 可使用 `flash_attention_2` (启用时的默认选项)，或设置为字符串 `"flash_attention_3"` 以使用 FlashAttention 3。

[gemini-code-assist]

The logic is correct, but it could be made slightly more explicit and readable by setting a default and then overriding it. This can make the intention clearer to future readers of the code.

Suggested change

	if self.use_flash_attention_kernel == "flash_attention_3":
	attn_impl = "flash_attention_3"
	else:
	attn_impl = "flash_attention_2"
	attn_impl = "flash_attention_2" # Default to flash_attention_2 when enabled
	if self.use_flash_attention_kernel == "flash_attention_3":
	attn_impl = "flash_attention_3"