feat(attention): add RoPE offset support for batch prefill

[MengAiDev]

📌 Description

• Add optional q_rope_offset and k_rope_offset parameters to batch prefill functions
• Update BatchAttention class to support RoPE offsets in batched attention
• Modify batch prefill CUDA kernels to incorporate RoPE offset parameters

🔍 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.

[gemini-code-assist]

Summary of Changes

Hello @MengAiDev, 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 introduces support for RoPE (Rotary Positional Embedding) offsets in batch prefill operations within the FlashInfer library. This enhancement allows for more flexible and advanced positional encoding handling during batched attention computations, particularly useful for models that utilize RoPE with dynamic or shifted positional information.

Highlights

• Extended Batch Prefill Functions: I've added optional q_rope_offset and k_rope_offset parameters to the core batch prefill functions (BatchPrefillWithRaggedKVCacheRun and BatchPrefillWithPagedKVCacheRun) in the CUDA backend.
• CUDA Kernel Integration: The underlying CUDA kernels have been updated to correctly process and utilize these new RoPE offset parameters, ensuring that the positional encoding is applied as intended during the attention calculation.
• Python API Exposure: The BatchAttention class in the Python API now exposes q_rope_offset and k_rope_offset as optional arguments to its run method, allowing users to easily pass these offsets from their Python code.
• Updated Pybind11 Bindings: The Pybind11 bindings for the batch prefill operations (ragged_run and paged_run) have been refactored to use lambda functions, which facilitates the seamless passing of the new optional RoPE offset tensors from Python to the C++/CUDA backend.

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 introduces RoPE offset support for batch prefill operations by adding optional q_rope_offset and k_rope_offset parameters to the CUDA kernels and their Python bindings. The changes are correctly propagated through the C++ layers. However, I've identified a critical issue in the Python BatchAttention class where an unsupported k_rope_offset parameter is passed to the paged attention kernel, which would lead to a runtime error. My feedback focuses on correcting this issue.

[gemini-code-assist]

The k_rope_offset parameter is not supported by the paged attention kernel that BatchAttention wraps. The underlying C++ function BatchPrefillWithPagedKVCacheRun only accepts q_rope_offset. Including k_rope_offset here will lead to a runtime error when calling the kernel. Please remove this parameter.