support DART for llava #403
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 introduces the DART token reduction method to the Llava model, allowing for sparsification and potentially improving inference efficiency. It includes a configuration file for specifying DART parameters and integrates the method into the model's forward pass using hooks. Additionally, it improves image token handling in other token reduction methods.
Highlights
- DART Token Reduction Method: Introduced the DART (Drop Add Recover Tokens) token reduction method for the Llava model, enabling sparsification by selectively pruning less important tokens.
- Configuration File: Added a configuration file (
dart.yml) to specify the parameters for the DART method, such as pruning location, reduction ratio, and pivot tokens. - Integration with Llava: Integrated DART into the Llava model by registering reduction modules and adding hooks to the forward pass to enable token pruning during inference.
- Image Token Handling: Enhanced image token handling in FastV and PyramidDrop by incorporating
IMAGE_TOKEN_INDEXfromllava.constants.
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 introduces support for the DART token reduction method for Llava models. This includes a new configuration file, the DART implementation, and its integration into the existing framework. The changes also refactor how IMAGE_TOKEN_INDEX is handled in other token reduction methods (FastV, PyramidDrop) for better consistency by centralizing it in the model's pruning configuration.
| ) - set(image_indices) | ||
|
|
||
| valid_indices_list = list(valid_indices) | ||
| for item in list(indices_set): |
There was a problem hiding this comment.
The list valid_indices_list can become empty inside this loop. If it does, valid_vectors on line 225 will be an empty tensor, and cos_sim.topk(...) on line 231 will raise an error. Please add a check at the beginning of the loop to handle this case, for example by breaking out of the loop if valid_indices_list is empty.
| valid_vectors, | ||
| dim=-1 | ||
| ) | ||
| top_k_indices = cos_sim.topk(TOKEN_TOPK).indices |
There was a problem hiding this comment.
The topk method will raise an error if TOKEN_TOPK is greater than the number of elements in cos_sim. The number of elements in cos_sim depends on len(valid_indices_list), which can be smaller than TOKEN_TOPK. You should ensure that the value passed to topk is not larger than the number of elements in the tensor.
| top_k_indices = cos_sim.topk(TOKEN_TOPK).indices | |
| top_k_indices = cos_sim.topk(min(TOKEN_TOPK, cos_sim.shape[0])).indices |
| reduction_ratio: 0.778 | ||
| max_num_trunction: 128 | ||
| pivot_image_token: 4 | ||
| pivot_text_token : 4 |
There was a problem hiding this comment.
There is an extra space before the colon. Please remove it for consistency.
pivot_text_token: 4
No description provided.