update fastervlm #398
update fastervlm #398
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 @SmudgedWings, 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 primarily focuses on refining the fastervlm.py module by improving variable naming conventions for better clarity and consistency. It also includes a minor code cleanup by removing commented-out sections and redundant assignments, contributing to a more maintainable and readable codebase.
Highlights
- Variable Renaming for Consistency: I've renamed
index_masktoindex_masksandimage_masktoimage_masksacross several functions (pruning_hook,get_image_mask_hook,prepare_inputs_for_llm_hook,prepare_inputs_hook). This change improves naming consistency, especially since these variables often represent collections of indices or masks. - Code Cleanup and Refinement: I've removed commented-out code related to alternative visual token pruning strategies (e.g., random scores) and a redundant assignment of
image_attentionstopruning_paras. This helps to clean up the codebase and remove dead or unused code. - Minor Shape Derivation Adjustment: In
pruning_hook, I've adjusted howB, N, Care derived fromimage_features.shape. WhileNstill represents the number of tokens, this change provides a more explicit way to capture all dimensions.
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.
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
The pull request updates the fastervlm.py file, primarily by refactoring variable names for clarity (e.g., index_mask to index_masks) and removing commented-out code. It also changes the dimensionality assumption for image_features and removes the sorting of token_indices and the storage of processed image_attentions. While the renaming improves maintainability, the changes to array shape assumptions and the removal of sorting/storage could introduce subtle issues if not fully accounted for in downstream logic.
| image_attentions = image_attentions.mean(dim=1) # (B, N) = (1, 576) | ||
|
|
||
| B, N = image_features.shape[:2] | ||
| B, N, C = image_features.shape |
There was a problem hiding this comment.
Changing image_features.shape[:2] to image_features.shape implies that image_features is now strictly expected to be a 3-dimensional tensor. If image_features could sometimes be 2-dimensional, this change would lead to a ValueError during unpacking. Please confirm that image_features will always have 3 dimensions (B, N, C) at this point in the execution.
| image_attentions = image_attentions.mean(dim=1) # (B, N) | ||
| token_indices = torch.topk(image_attentions, k=visual_token_num, dim=1)[1] # (B, T) | ||
| token_indices = torch.sort(token_indices, dim=1)[0] # (B, T) | ||
|
|
There was a problem hiding this comment.
The removal of torch.sort(token_indices, dim=1)[0] means that the token_indices will no longer be sorted. While the scatter_ operation itself does not strictly require sorted indices, if any downstream logic implicitly relied on these indices being in a specific order (e.g., for consistency or debugging), this change could introduce subtle behavioral differences. Please ensure that the order of token_indices does not matter for any subsequent operations.
| pruning_paras['index_masks'] = index_masks | ||
|
|
There was a problem hiding this comment.
The processed image_attentions (after mean(dim=1)) was previously stored back into pruning_paras['image_attentions']. This line has been removed. If any other part of the system or subsequent hooks rely on pruning_paras['image_attentions'] containing the mean-reduced attention, this could lead to incorrect behavior. If this processed value is not used elsewhere, consider adding a comment to clarify its scope.
2 participants
No description provided.