Fix one missing arg in DeepEP

[ch-wan]

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 @ch-wan, 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 here, providing a summary of this pull request to help everyone get up to speed quickly. This PR, titled "Fix one missing arg in DeepEP", addresses an issue where the DeepEPMoELayer class was not correctly passing the num_fused_shared_experts argument to its parent class, ExpertParallelMoELayer. Although num_fused_shared_experts is not currently supported in the Expert Parallel (EP) mode used by DeepEPMoELayer (and an assertion is added to enforce this), the parent constructor still expects the argument. This change ensures the argument is correctly included in the constructor call, resolving the missing argument issue and improving code clarity by switching the parent constructor call to use keyword arguments.

Highlights

• Argument Fix: The primary change is adding the num_fused_shared_experts argument to the DeepEPMoELayer constructor and passing it correctly to the ExpertParallelMoELayer parent class constructor.
• Constraint Enforcement: An assertion is added in the ExpertParallelMoELayer constructor to explicitly state and enforce that num_fused_shared_experts must be 0 in this configuration, clarifying that fused shared experts are not supported in EP mode.
• Code Cleanup: Internal handling and passing of num_fused_shared_experts within the ExpertParallelMoELayer class (where it's not used) are removed for clarity, and the super().__init__ call in DeepEPMoELayer is updated to use keyword arguments.

Changelog

• python/sglang/srt/layers/moe/ep_moe/layer.py
  • Added an assertion in ExpertParallelMoELayer.__init__ to ensure num_fused_shared_experts is 0, indicating it's not supported in EP mode.
  • Removed the storage of num_fused_shared_experts as a class member in ExpertParallelMoELayer.
  • Removed the passing of num_fused_shared_experts to the moe_align_block function call within ExpertParallelMoELayer.forward.
  • Added num_fused_shared_experts with a default value of 0 to the DeepEPMoELayer.__init__ method signature.
  • Passed the num_fused_shared_experts argument to the super().__init__ call in DeepEPMoELayer.__init__.
  • Switched the super().__init__ call in DeepEPMoELayer.__init__ to use keyword arguments for improved readability.

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.

---

A missing argument,
A small bug, quickly sent.
Code now compiles,
With happy smiles,
Development's intent.

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 PR correctly addresses a critical bug in DeepEPMoE.__init__ where arguments to the superclass constructor EPMoE.__init__ were misaligned due to a missing argument. The fix involves correctly adding num_fused_shared_experts to the DeepEPMoE constructor and passing it to the superclass.

The related changes in EPMoE (adding an assertion for num_fused_shared_experts == 0 and removing its subsequent usage within that class if it's always 0) are consistent and improve clarity about its current capabilities.

Summary of Findings

• Critical Bug Fix: Argument Misalignment in Constructor: The primary change correctly passes num_fused_shared_experts to the EPMoE superclass constructor within DeepEPMoE.__init__. This resolves a critical bug where arguments were previously misaligned. The explicit use of keyword arguments in the super().__init__ call further enhances the robustness and readability of this fix.
• Suggestion: Unit Testing for Constructor Logic: To ensure the correctness of the fix and prevent future regressions, it's recommended to add unit tests. These tests should specifically cover the initialization of DeepEPMoE and verify that arguments, particularly num_fused_shared_experts, are correctly passed to and handled by the EPMoE superclass.

Merge Readiness

This PR addresses a critical bug concerning argument passing in class initialization, and the fix appears correct and robust. The use of keyword arguments is a good practice. To increase confidence and prevent regressions, adding unit tests for the constructor logic is recommended. As an AI, I cannot approve pull requests; please ensure another human reviewer approves this change, especially after considering the test suggestion.

[gemini-code-assist]

This change correctly fixes a critical bug. Previously, num_fused_shared_experts was not explicitly passed to the EPMoE superclass constructor, and due to the use of positional arguments, this led to a misalignment of subsequent arguments (e.g., quant_config from DeepEPMoE would have been passed as num_fused_shared_experts to EPMoE). This could cause incorrect behavior or runtime errors, especially with the new assertion in EPMoE that num_fused_shared_experts must be 0.

Passing num_fused_shared_experts explicitly now resolves this critical issue.

Additionally, the switch from positional arguments to keyword arguments for the super().__init__ call is an excellent improvement for code readability and maintainability. It makes the call more robust to future changes in the superclass's __init__ signature.