fix scale_shift kernel for wan2.2-5b #828
Conversation
Summary of ChangesHello @gushiqiao, 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 introduces a crucial adjustment to the 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. 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
The pull request aims to fix the scale_shift kernel for a specific model version where scale and shift tensors have a shape of [L, 1, C]. However, the condition used to detect this case is too broad and can incorrectly process other valid tensor shapes, leading to runtime errors. I've suggested a more specific condition to correctly target the intended shape without affecting other cases.
| if scale.shape[1] == 1: | ||
| scale = scale.squeeze(1).unsqueeze(0) | ||
| if shift.shape[1] == 1: | ||
| shift = shift.squeeze(1).unsqueeze(0) |
There was a problem hiding this comment.
The condition scale.shape[1] == 1 is too general and can cause issues. For example, if scale has a shape of [B, 1, C], which is a valid shape handled by the existing logic, this condition will be true. The tensor will be incorrectly reshaped from [B, 1, C] to [1, B, C], which will later cause a runtime error during the expand operation if B > 1.
To fix this, the condition should be more specific to target only the [L, 1, C] shape as mentioned in the comment.
| if scale.shape[1] == 1: | |
| scale = scale.squeeze(1).unsqueeze(0) | |
| if shift.shape[1] == 1: | |
| shift = shift.squeeze(1).unsqueeze(0) | |
| if scale.dim() == 3 and scale.shape[0] == L and scale.shape[1] == 1: | |
| scale = scale.squeeze(1).unsqueeze(0) | |
| if shift.dim() == 3 and shift.shape[0] == L and shift.shape[1] == 1: | |
| shift = shift.squeeze(1).unsqueeze(0) |
No description provided.