[bugfix] fix PeftModel save_pretrained selected_adapters when using custom adapter_name#8337
[bugfix] fix PeftModel save_pretrained selected_adapters when using custom adapter_name#8337alphadl wants to merge 1 commit intomodelscope:mainfrom
Conversation
Summary of ChangesHello, 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 Swift framework's model saving mechanism, specifically impacting PeftModel instances configured with custom adapter names. Previously, attempts to save such models would result in a ValueError because the save_pretrained method was hardcoded to look for a 'default' adapter. The changes introduce a robust way to identify the currently active adapter(s) and ensure these are correctly passed during the saving process, thereby preventing errors and allowing seamless saving of models with custom PEFT configurations. Highlights
🧠 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. Changelog
Activity
Using Gemini Code AssistThe 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
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 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
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request addresses a bug where saving a PeftModel with a custom adapter name would fail. The fix correctly introduces a helper function, _get_peft_selected_adapters, to determine the active adapters and uses this function when calling save_pretrained. The changes are logical and are accompanied by new unit tests. I have identified a potential issue in the implementation of the new helper function that could affect multi-adapter scenarios and have provided a suggestion for a more robust implementation.
| we must pass the same name(s) to save_pretrained; otherwise PEFT raises because | ||
| supported adapter names do not include 'default'. See: gh#8336. | ||
| """ | ||
| active = getattr(model, 'active_adapter', None) or getattr(model, 'active_adapters', None) |
There was a problem hiding this comment.
The current logic for determining active adapters might be incorrect in a multi-adapter scenario. active_adapter is a property that returns the first adapter from active_adapters. By checking active_adapter first, you might only get the first active adapter instead of all of them if multiple are active.
To correctly retrieve all active adapters, you should check for active_adapters first.
| active = getattr(model, 'active_adapter', None) or getattr(model, 'active_adapters', None) | |
| active = getattr(model, 'active_adapters', None) or getattr(model, 'active_adapter', None) |
|
Please take a look when you have time. This fixes the PeftModel save error when using a custom |
1 participant
[bugfix] fix PeftModel save_pretrained selected_adapters when using custom adapter_name
#8336
When users pass a custom
adapter_name(e.g.vision_only_lora) toSwift.prepare_model(), checkpoint save was still passingselected_adapters=['default']toPeftModel.save_pretrained(), causing:Changes:
swift/trainers/mixin.py: Add_get_peft_selected_adapters(model)to resolveactive_adapter/active_adaptersfrom the PeftModel; use it in_save_model()for both PeftModel save paths instead of hardcoded['default'].tests/tuners/test_peft.py: Add tests for default adapter, custom adapter_name, and mock (None/str/list).