[DSv32] Add returning DSA topk indices

[zianglih]

Motivation

#16856

Modifications

Incrementally add DSA topk indices capturing on top of #12162.

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.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello @zianglih, 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 enhances the SGLang framework by introducing the ability to capture and return Dynamic Sparse Attention (DSA) top-k indices. This feature allows users to explicitly request these indices during generation, providing deeper insights into the model's attention mechanism. The changes involve extending the request and output data structures, implementing a dedicated capturing and buffering system for these indices, and integrating their collection within the model's forward pass and the overall request processing pipeline.

Highlights

• API Extension: Introduced a new return_dsa_topk_indices flag in request input structures (GenerateReqInput, TokenizedGenerateReqInput) to allow users to explicitly request Dynamic Sparse Attention (DSA) top-k indices.
• Data Structure Enhancement: Extended output data structures (BatchTokenIDOutput, BatchStrOutput) to include output_dsa_topk_indices, enabling the return of these captured indices.
• DSA Index Capturing Mechanism: Implemented a dedicated system within RoutedExpertsCapturer to allocate and manage device and host buffers for DSA topk indices, including new methods for capturing and synchronizing this data.
• Model Integration: Integrated the DSA topk index capturing into the deepseek_v2 model's forward_absorb_prepare method, ensuring these indices are collected during the model's execution.
• System-wide Propagation: Ensured the return_dsa_topk_indices flag and the captured data are correctly propagated and processed through the scheduling, detokenization, and output processing managers.
• Configuration and Benchmarking Support: Added enable_return_dsa_topk_indices as a server argument for global control and included support for this feature in the benchmarking utility (bench_serving.py).

🧠 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 adds the capability to return DSA topk indices by introducing a return_dsa_topk_indices flag. The changes are mostly additive, plumbing the new flag and data structures through various components from the entrypoint to the model execution layer. My review identifies a critical bug where a parameter shadowing issue prevents the new feature from working correctly, and I've provided a code suggestion to fix it. I also have a medium-severity suggestion to refactor duplicated code for better maintainability.