Skip to content

[BugFix][VLM] Correct weight loading with tie_word_embeddings == False#15398

Merged
merrymercy merged 4 commits intosgl-project:mainfrom
antgroup:py/fix-weights
Dec 27, 2025
Merged

[BugFix][VLM] Correct weight loading with tie_word_embeddings == False#15398
merrymercy merged 4 commits intosgl-project:mainfrom
antgroup:py/fix-weights

Conversation

@ZhengWG
Copy link
Copy Markdown
Collaborator

@ZhengWG ZhengWG commented Dec 18, 2025

Motivation

Cannot reload embed_tokens.weight into lm_head.weight when tie_word_embeddings is False

Modifications

Accuracy Tests

Benchmarking and Profiling

Checklist

@gemini-code-assist
Copy link
Copy Markdown
Contributor

gemini-code-assist Bot commented Dec 18, 2025

Summary of Changes

Hello @ZhengWG, 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 addresses a critical bug in the VLM's weight loading mechanism. Previously, the system would implicitly tie the input embedding weights to the language model head weights during loading, regardless of the model's configuration. The fix ensures that this weight-tying behavior is now correctly conditional, preventing unintended parameter sharing and maintaining model integrity when tie_word_embeddings is disabled.

Highlights

  • Bug Fix: Incorrect Weight Loading: Resolved an issue where the embed_tokens.weight was being incorrectly reloaded into lm_head.weight even when the tie_word_embeddings configuration was set to False.
  • Conditional Weight Tying Logic: Introduced a conditional check within the load_weights function to ensure that the special handling for model.embed_tokens.weight (which affects lm_head.weight) is only applied when self.config.tie_word_embeddings is explicitly True.

🧠 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[bot]
gemini-code-assist Bot reviewed Dec 18, 2025
View reviewed changes
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request fixes a bug in the weight loading logic for Qwen2_5_VLForConditionalGeneration. The original code incorrectly tied the lm_head weights to the embed_tokens weights on the last pipeline parallel rank, regardless of the tie_word_embeddings configuration. The added check for self.config.tie_word_embeddings ensures this weight tying only happens when it's intended, which is the correct behavior. The change is accurate and necessary.

@ZhengWG
Copy link
Copy Markdown
Collaborator Author

ZhengWG commented Dec 18, 2025

@XucSh Could you help review the logic? Since this issue was introduced by the changes in #15138.

@XucSh
Copy link
Copy Markdown
Collaborator

XucSh commented Dec 18, 2025

@XucSh Could you help review the logic? Since this issue was introduced by the changes in #15138.

Thanks for your fix.

@ZhengWG
Copy link
Copy Markdown
Collaborator Author

ZhengWG commented Dec 22, 2025

@XucSh Can you help trigger CI?

@yuan-luo
Copy link
Copy Markdown
Collaborator

yuan-luo commented Dec 22, 2025

/rerun-failed-ci

1 similar comment
@yuan-luo
Copy link
Copy Markdown
Collaborator

yuan-luo commented Dec 22, 2025

/rerun-failed-ci

@XucSh
Copy link
Copy Markdown
Collaborator

XucSh commented Dec 23, 2025

/tag-and-rerun-ci

@yuan-luo
Copy link
Copy Markdown
Collaborator

yuan-luo commented Dec 23, 2025

/rerun-failed-ci

@yuan-luo
Copy link
Copy Markdown
Collaborator

yuan-luo commented Dec 26, 2025

/rerun-failed-ci

@yuan-luo yuan-luo mentioned this pull request Dec 27, 2025
Merged
6 tasks
@yuan-luo yuan-luo self-requested a review December 27, 2025 00:36
@merrymercy merrymercy merged commit cd3289c into sgl-project:main Dec 27, 2025
144 of 150 checks passed
YChange01 pushed a commit to YChange01/sglang that referenced this pull request Jan 13, 2026
@ZhengWG @YChange01
[BugFix][VLM] Correct weight loading with tie_word_embeddings == False (
e20e781 
sgl-project#15398)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@yuan-luo yuan-luo yuan-luo approved these changes

+2 more reviewers

@gemini-code-assist gemini-code-assist[bot] gemini-code-assist[bot] left review comments

@XucSh XucSh XucSh approved these changes

Reviewers whose approvals may not affect merge requirements

Assignees

No one assigned

Labels

run-ci

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@ZhengWG @XucSh @yuan-luo @merrymercy