[1/N][Sparse With Hicache]: Add Sparse Interface

[hzh0425]

Motivation

This PR primarily introduces a generic Retrievable Sparse Algorithm interface.
Upstream PR：#14619

Modifications

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.
• Work with maintainers to merge your PR. See the PR Merge Process

[gemini-code-assist]

Summary of Changes

Hello @hzh0425, 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 establishes a foundational interface for integrating diverse sparse attention algorithms within the system. By introducing a BaseSparseAlgorithm and SparseMode enum, it provides a standardized way to implement and manage KV cache compression techniques. The PR also includes two initial algorithm implementations, DeepSeekNSAAlgorithm and KnormPageAlgorithm, demonstrating the extensibility of this new framework. This change is crucial for enabling more efficient memory usage and potentially faster inference by selectively attending to important KV cache entries.

Highlights

• Introduced BaseSparseAlgorithm: A new abstract base class for defining various retrievable KV cache compression algorithms, supporting both page-wise and token-wise sparsity.
• Defined SparseMode Enum: An enumeration to categorize sparse attention granularity modes, including PAGE_WISE, TOKEN_WISE, and DEEPSEEK_TOKEN_WISE.
• Added DeepSeekNSAAlgorithm: A concrete implementation for handling DeepSeek NSA specific sparse attention logic.
• Implemented KnormPageAlgorithm: An experimental page-wise sparse attention algorithm utilizing ChunkKV-style scoring based on key L2 norms.
• Included Unit Tests: Comprehensive unit tests for the KnormPageAlgorithm to ensure its correctness and functionality.

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 introduces a valuable abstraction for sparse attention algorithms by adding a generic BaseSparseAlgorithm interface. The changes include the base class, two initial implementations (DeepSeekNSAAlgorithm and KnormPageAlgorithm), and unit tests for one of them. My review identifies a critical issue in DeepSeekNSAAlgorithm that would lead to a runtime error due to an incorrect return value. I've also pointed out several medium-severity issues in KnormPageAlgorithm related to hardcoded values, dead code, incorrect type hints, and verbose logging that could impact maintainability and performance. Overall, this is a solid architectural improvement, and addressing these points will enhance the robustness and quality of the new sparse attention framework.

[hzh0425]

we performed a refactoring by removing the distinction between PageWise and TokenWise, as TokenWise is essentially a special case of PageWise with PageSize = 1.

In base_algorithm.py, I implemented a BaseSparseAlgorithmImpl as a common base class. Subclasses can integrate into the pipeline by implementing a few key methods, and they also have the flexibility to override critical methods such as retrieve_topk and update_representations to customize their specific logic.

Additionally, we plan to improve the performance of BaseSparseAlgorithmImpl in the next PR.
This PR focuses on defining the foundational interfaces and Implement a reasonable sparse Algo.

Could you please review it again when you have time? @xiezhq-hermann

The Accuracy Test for Quest Sparse Algo(Implemented by @magicYang1573):

GSM8K

llama 8B Test: (The accuracy (ACC) of the original model is 0.94.)

CompressionRate	0.3	0.5	0.7
Accuracy ACC	0.92	0.92	0.92

aime25

Qwen32B Test:

Original Mode：

Quest Sparse Mode with 0.5 compresstion rate

[hzh0425]

/rerun-failed-ci