Fix/moe loss #5316
Fix/moe loss #5316
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @tastelikefeet, 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!
I've addressed a bug related to Mixture of Experts (MoE) loss calculation within our sequence parallel training framework. This ensures that the auxiliary loss from MoE models, specifically derived from router_logits, is correctly gathered and incorporated into the total loss. This fix improves the training stability and accuracy for MoE models when trained with sequence parallelism.
Highlights
- Loss Gathering Utility: I've introduced a new
GatherLossutility and integrated it into the sequence parallel training setup to correctly aggregate loss components, specificallyrouter_logits, across distributed devices. - MoE Router Logits Handling: I've implemented a forward hook (
moe_aux_loss_hook) that specifically handles therouter_logitsgenerated by Mixture of Experts (MoE) models. This hook ensures that these logits are properly gathered and reshaped across sequence parallel ranks before auxiliary loss computation. - MoE Auxiliary Loss Integration: I've updated the main
compute_lossfunction to correctly incorporate the auxiliary loss from MoE models. This loss is now added to the primary loss, weighted by a configurable coefficient, ensuring that the MoE routing mechanism is properly regularized during training.
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 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 or fill out our survey 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.
Footnotes
-
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. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces a fix for handling Mixture-of-Experts (MoE) auxiliary loss, particularly in a sequence parallel training setup. The changes are well-structured and address the issue effectively across multiple files.
In swift/trainers/sequence_parallel/utils.py, the GatherLoss function has been generalized to handle cases where labels are not provided, which is a necessary change for processing MoE router logits.
In swift/trainers/sequence_parallel/ulysses.py, a new forward hook moe_aux_loss_hook is added to correctly gather router_logits from all sequence parallel workers. This ensures the auxiliary loss is computed on the full, un-sharded sequence, which is crucial for correctness.
Finally, in swift/trainers/trainers.py, the main compute_loss function is updated to correctly incorporate the MoE auxiliary loss into the total loss.
Overall, the changes are logical, consistent, and appear to be a solid fix for the problem. I have not found any issues with the implementation.
PR type
PR information
Write the detail information belongs to this PR.
Experiment results
Paste your experiment result here(if needed).