feat: Support sliding window for persistent kernel

[Edenzzzz]

📌 Description

Also changed kv chunk size from an array to a single element.
Precision still seems buggy, will try to fix tmr

🔍 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!

I've implemented new features to enhance the persistent kernel's attention mechanism. This includes adding support for a sliding window, which optimizes performance and memory usage for long sequences by limiting the attention scope. Additionally, I've introduced logits soft capping, a technique that can improve the stability and quality of attention calculations. These changes are integrated into the core attention logic, exposed through the Python API, and thoroughly tested to ensure correctness and efficiency.

Highlights

• Sliding Window Attention: I've introduced support for sliding window attention, allowing the attention mechanism to focus on a limited, local context rather than the entire sequence history. This is crucial for managing memory and computation with very long sequences.
• Logits Soft Capping: I've added a logits soft capping feature, which applies a tanh transformation to the attention logits. This can help stabilize training and improve model performance, especially in models like Gemini, Grok, and Gemma-2 that utilize this technique.
• Kernel and JIT Integration: The underlying CUDA kernels and JIT compilation system have been updated to dynamically incorporate these new features, ensuring efficient execution based on whether sliding window or logits soft capping are enabled.
• Increased Scheduler Capacity: I've expanded the capacity limits for internal planning buffers in the attention scheduler, increasing max_total_num_works and max_num_kv_splits to accommodate larger and more complex attention workloads.
• Enhanced Test Coverage: Comprehensive unit tests have been added and parameterized to validate the correctness of both the sliding window and logits soft capping functionalities across various configurations.

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 support for sliding window attention and logits soft capping in the persistent kernel. The changes span across the Python API, JIT compilation, and CUDA kernels. The implementation looks mostly correct, but I've found a critical bug related to a type mismatch for the window_left parameter that would cause the sliding window logic to fail. I've also pointed out a missing validation check in the Python API and a minor code style improvement in the CUDA kernel template.

[gemini-code-assist]

The member variable window_left is only used within this mask generation logic. It's better to declare it as a local const variable to improve code clarity and prevent accidental misuse elsewhere. This also removes the need for the member declaration at line 30.

    if constexpr (use_sliding_window) {
      uint32_t qo_len = params.q_len[work_idx];
      uint32_t kv_len = params.kv_len[work_idx];
      const auto window_left_val = (params.window_left >= 0) ? static_cast<uint32_t>(params.window_left) : kv_len;
      mask &= (kv_idx + qo_len + window_left_val >= kv_len + qo_idx);
    }

[Edenzzzz]

Only wrong precision when split kv

[yzh119]

Sliding window is a little bit tricky with split-kv in terms of boundary computation.

[Edenzzzz]

Even when I set window_tile_idx = kv_tile_idx, there's still a bit of mismatch in the second case, so other indices could also be problematic