[3/N][Sparse With Hicache]: Init sparse coordinator

[hzh0425]

Motivation

This PR introduces a flexible configuration framework for hierarchical sparse attention with retrievable KV cache compression algorithms (Quest, PQCache, SnapKV, etc.) in SGLang.
Upstream PR：#14619

Modifications

Coordinator Lifecycle Framework

1. Designed for decode-phase retrievable algorithms
2. Request lifecycle hooks: https://github.com/sgl-project/sglang/pull/16086/files#diff-cc98e3ecfeb5c89804f818dd8c4be3b3b15ef803d5555eb628439018d8e092cfR65
   • on_request_begin/end: State management
   • forward_begin/end: KVCache offloading synchronization
   • attention_begin/end: Sparse retrieval and representation updates

Attention BackendAdaptor: https://github.com/sgl-project/sglang/pull/16086/files#diff-e0e2432b5521042dbac9e254d142233b2d8366b964ae83ce96f748730db53276R25

• BackendAdaptor transforms logical sparse indices to backend-specific metadata
• Supports FlashAttention (page table modification) now
• Extensible for future backends

Accuracy Tests

#14741 (comment)

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 framework for hierarchical sparse attention within SGLang, enabling the integration of various retrievable KV cache compression algorithms. The core SparseCoordinator orchestrates the entire sparse attention process, from request initiation to KV cache management and attention metadata adaptation. This design promotes modularity and extensibility, allowing for future support of different sparse algorithms and attention backends, starting with FlashAttention. The changes also introduce new configuration options for users to enable and customize sparse attention behavior.

Highlights

• Sparse Coordinator Framework: Introduced a new SparseCoordinator to manage the lifecycle of hierarchical sparse attention, particularly for decode-phase retrievable algorithms.
• Request Lifecycle Hooks: Implemented specific hooks (on_request_begin/end, forward_begin/end, attention_begin/end) within the coordinator for state management, KV cache offloading synchronization, and sparse retrieval/representation updates.
• Attention Backend Adaptor: Designed an abstract BackendAdaptor to translate logical sparse indices into backend-specific attention metadata, with an initial implementation for FlashAttentionAdaptor.
• Configurability: Added new command-line arguments (--enable-hierarchical-sparse-attention, --hierarchical-sparse-attention-extra-config) to ServerArgs for flexible configuration of sparse attention algorithms and backends.

🧠 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 introduces a new framework for hierarchical sparse attention, including a coordinator, algorithms, and backend adaptors. The changes are well-structured, separating concerns into distinct modules for better maintainability and extensibility. The configuration parsing is robust, handling JSON input for extra settings and providing sensible defaults. The integration with existing ServerArgs is also clear, adding new command-line arguments for enabling and configuring sparse attention.

[hzh0425]

@xiezhq-hermann could you take a review

[wqlxx]

@hzh0425 This pr support Speculative Decoding ?

[xiezhq-hermann]

Is there a dependency between this one and the last one (#15807)?

[hzh0425]

Is there a dependency between this one and the last one (#15807)?

It is independent and has no dependencies.