fix bugs for dart fasyv sparsevlm and update ci #412
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 addressing several bugs and making improvements to the DART, FastV, and SparseVLM token reduction algorithms. It also includes necessary updates to the continuous integration (CI) configurations to reflect changes in repository paths. The core changes involve refining token indexing, optimizing attention mechanism parameter handling, and introducing a new configurable flag for token merging in SparseVLM, all aimed at enhancing the robustness and configurability of these compression methods.
Highlights
- CI Configuration Updates: Updated hardcoded repository paths in
awq_w4a16_fakequant_eval.ymlandgptq_w_only.ymlfrom/home/runner/work/llmc/llmc/to/home/runner/work/LightCompress/LightCompress/to reflect a potential repository rename or move. - Token Indexing Bug Fixes: Addressed an indexing bug in
llmc/compression/token_reduction/dart.pyandllmc/compression/token_reduction/fastv.pyby changingtorch.where(...)[0].item()totorch.where(...)[0][0].item()to ensure the correct first element is retrieved. - SparseVLM Logic Refinements: Introduced a new
merge_flagconfiguration option inconfigs/sparsification/methods/SparseVLM/sparsevlm.ymland integrated it into theSparseVLMalgorithm to conditionally control token merging and clustering. This also includes refactoring ofpast_key_valueand attention cache handling for improved correctness and efficiency. - Llava Model Integration Improvements: Implemented new hooks (
update_kwargs_hookandget_attn_logits_hook) inllmc/compression/token_reduction/sparsevlm.pyto better manage and update attention-related parameters (likeposition_ids,cache_position,position_embeddings) during the forward pass for theLlavamodel.
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
This pull request includes bug fixes for DART, FastV, and SparseVLM token reduction methods, along with CI configuration updates. The changes primarily address index selection and internal state management. Review focuses on error handling, thread safety, and code maintainability.
| input_ids[0][attention_mask[0]] == pruning_paras['IMAGE_TOKEN_INDEX'] | ||
| ) | ||
| pruning_paras['image_token_start_index'] = torch.where(token_indices)[0].item() | ||
| pruning_paras['image_token_start_index'] = torch.where(token_indices)[0][0].item() |
There was a problem hiding this comment.
Ensure that token_indices always contains at least one element before accessing token_indices[0][0]. Otherwise, this could raise an IndexError if no image tokens are found.
pruning_paras['image_token_start_index'] = token_indices[0][0].item() if len(token_indices[0]) > 0 else 0| token_indices = \ | ||
| input_ids[0][attention_mask[0]] == pruning_paras['IMAGE_TOKEN_INDEX'] | ||
| pruning_paras['image_token_start_index'] = torch.where(token_indices)[0].item() | ||
| pruning_paras['image_token_start_index'] = torch.where(token_indices)[0][0].item() |
There was a problem hiding this comment.
Ensure that token_indices always contains at least one element before accessing token_indices[0][0]. Otherwise, this could raise an IndexError if no image tokens are found.
pruning_paras['image_token_start_index'] = token_indices[0][0].item() if len(token_indices[0]) > 0 else 0| from .token_reduction_module import TokenReductionModule | ||
| from .utils import prefill_wrapper, prefill_wrapper_model | ||
|
|
||
| layer_dict = {} |
There was a problem hiding this comment.
Introducing a global layer_dict can lead to unexpected behavior if multiple SparseVLM instances are used concurrently. Refactor to use an instance variable to ensure thread safety and proper encapsulation.
| model: | ||
| type: Opt | ||
| path: /home/runner/work/llmc/llmc/ci_check/opt-125m | ||
| path: /home/runner/work/LightCompress/LightCompress/ci_check/opt-125m |
There was a problem hiding this comment.
The file path is hardcoded. Consider using an environment variable that can be substituted by the CI system to improve portability.
| model: | ||
| type: Opt | ||
| path: /home/runner/work/llmc/llmc/ci_check/opt-125m | ||
| path: /home/runner/work/LightCompress/LightCompress/ci_check/opt-125m |
There was a problem hiding this comment.
The file path is hardcoded. Consider using an environment variable that can be substituted by the CI system to improve portability.
| pruning_loc: [2] # [2, 6, 15] | ||
| retained_tokens: 192 | ||
| init_token_total_shape: 668 | ||
| merge_flag: False |
There was a problem hiding this comment.
Consider the implications of setting merge_flag to False. Ensure this aligns with the intended behavior of disabling token merging, as it might impact performance or accuracy.
| t_token_idx = t_token_idx[1] + v_token_start + v_token_num | ||
| L, S = query_states.size(-2), key_states.size(-2) |
There was a problem hiding this comment.
This change removes parentheses around any_states.permute and any_states.reshape. Confirm that this change does not affect the order of operations or the intended result.
| self.blocks[block_idx].self_attn.register_forward_pre_hook( | ||
| functools.partial( | ||
| update_kwargs_hook, | ||
| pruning_pars=self.pruning_paras, | ||
| layer_idx=block_idx, | ||
| ), | ||
| with_kwargs=True | ||
| ) |
There was a problem hiding this comment.
Consider consolidating the duplicate registration of update_kwargs_hook and get_attn_logits_hook for Llava models to reduce code duplication and improve maintainability.
2 participants
No description provided.