Speed up rebalancing when using non-static dispatch algorithms

[fzyzcjy]

Motivation

Modifications

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.

[gemini-code-assist]

Hello @fzyzcjy, 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!

Summary of Changes

Hello! Gemini or gemini-code-assist here, providing a summary of this pull request. This PR aims to improve performance, specifically the rebalancing process, when using non-static expert dispatch algorithms. It achieves this by making the logical_to_rank_dispatch_physical_map optional and only computing and storing it when the 'static' dispatch algorithm is configured. This avoids unnecessary computation and memory usage for algorithms like 'random' dispatch where this map is not needed.

Highlights

• Performance Optimization: The core goal is to speed up rebalancing by avoiding unnecessary work when using non-static dispatch algorithms.
• Conditional Data Storage: The logical_to_rank_dispatch_physical_map is now conditionally computed and stored only if the dispatch algorithm is 'static'. For other algorithms, it is set to None.
• Optional Fields: Related fields (logical_to_rank_dispatch_physical_map and partial_logical_to_rank_dispatch_physical_map) have been updated to be Optional[torch.Tensor] to reflect that they may not always be present.
• Refactored Initialization and Update: The initialization (_init_raw, init_by_mapping, init_by_eplb, init_new) and update logic have been adjusted to correctly handle the new optional nature of these fields.

Changelog

• python/sglang/srt/managers/expert_location.py
  • Made logical_to_rank_dispatch_physical_map an optional torch.Tensor.
  • Removed assertions related to the shape of logical_to_rank_dispatch_physical_map in __post_init__.
  • Added server_args parameter to _init_raw, init_by_mapping, and init_by_eplb methods.
  • Modified _init_raw to compute logical_to_rank_dispatch_physical_map only if server_args.ep_dispatch_algorithm is 'static', otherwise sets it to None.
  • Updated the update method to handle the optional logical_to_rank_dispatch_physical_map, ensuring both source and destination are either None or not None before attempting to copy.
• python/sglang/srt/managers/expert_location_dispatch.py
  • Made partial_logical_to_rank_dispatch_physical_map an optional torch.Tensor.
  • Updated the init_new method to conditionally assign the value to partial_logical_to_rank_dispatch_physical_map based on whether the source logical_to_rank_dispatch_physical_map is None.

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 is currently in preview and 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 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.

---

Data can be optional,
Like a tensor conditional.
If not static, it's gone,
Performance moves on,
Making code less emotional.

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 aims to speed up rebalancing when non-static dispatch algorithms are used by making the logical_to_rank_dispatch_physical_map optional and only computing it when the dispatch algorithm is 'static'. The changes are well-targeted and logically implement this optimization.

The introduction of Optional for logical_to_rank_dispatch_physical_map is handled consistently across ExpertLocationMetadata and ExpertLocationDispatchInfo, including updates to initialization, assertions, and the update method.

Overall, the code quality is good, and the changes are clear. I have one point for clarification regarding an assertion in the update method.

It would also be beneficial to fill out the PR description (Motivation, Modifications) and the checklist to provide more context for reviewers and for future reference.

Summary of Findings

• PR Description and Checklist: The pull request description (Motivation, Modifications) and the checklist are currently empty. Filling these out would provide valuable context for reviewers and for future documentation.

Merge Readiness

The changes in this PR appear to be well-implemented for the stated goal of speeding up rebalancing with non-static dispatch algorithms. There is one medium-severity point for clarification regarding an assertion in the update method. Once that is clarified, and assuming the author addresses the checklist and PR description, the PR should be in good shape. As an AI reviewer, I am not authorized to approve pull requests; please ensure further review and approval from team members before merging.

[gemini-code-assist]

The assertion (src is not None) == (dst is not None) ensures that self and other have the same optional status for the field being updated (e.g., logical_to_rank_dispatch_physical_map). This implies they were likely configured with the same ep_dispatch_algorithm.

Could you clarify if it's an expected invariant that the update method is only called between ExpertLocationMetadata instances that have a consistent configuration regarding the presence of this map? If there are scenarios where self and other might have different configurations (e.g., one static, one dynamic), this assertion would prevent the update. Understanding the intended state transitions here would be helpful.