[Attention] feat: support PrefixLM

[AlpinDale]

For moondream3 support, in a later PR.

[gemini-code-assist[bot]]

Summary of Changes

Hello @AlpinDale, I'm Gemini Code Assist[^1]! 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 lays the groundwork for supporting Prefix Language Modeling (PrefixLM) within the FlexAttention mechanism. It introduces the necessary masking logic and configuration parameters to enable models like moondream3 to utilize PrefixLM, which is a crucial step for their integration in a subsequent PR. The changes primarily involve adding a new mask type and integrating it into the attention metadata and builder processes.

Highlights

• PrefixLM Masking Function: Introduced a new prefixlm_mask_mod function that defines the specific masking logic for Prefix Language Modeling, allowing bidirectional attention within a prefix and causal attention for the suffix.
• FlexAttentionMetadata Update: The FlexAttentionMetadata dataclass has been extended with prefixlm (boolean) and prefix_len (integer) fields to control and specify PrefixLM behavior.
• Conditional Mask Application: Modified the FlexAttention class's build_block_mask and __post_init__ methods to dynamically select and apply the prefixlm_mask_mod when PrefixLM is enabled, alongside existing causal and bidirectional masking options.
• Builder Method Enhancement: The FlexAttentionBuilder.build method now accepts prefixlm and prefix_len as parameters, allowing for the configuration of PrefixLM during the attention metadata construction.
• Forward Pass Logic Adjustment: Updated the forward method in FlexAttention to incorporate the prefixlm flag when determining whether to apply non-causal attention, ensuring correct behavior for PrefixLM models.

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

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 :thumbsup: and :thumbsdown: 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.

[^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.

[AlpinDale]

/gemini review