support multi turn questions #420
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 @helloyongyang, 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 significantly enhances the LLaVA model's capabilities by adding robust support for multi-turn question answering. It ensures that the model can maintain conversational context and efficiently process subsequent questions by reusing image token pruning decisions from the initial turn. A new evaluation mode is also introduced to facilitate testing of these multi-turn interactions.
Highlights
- Multi-Turn Question Answering Support: Introduced core functionality to enable the LLaVA model to handle multi-turn conversations, allowing it to maintain context across successive questions related to the same image(s). This is crucial for interactive AI applications.
- Context-Aware Token Reduction: Modified the random token pruning mechanism to be aware of conversational turns. Image token pruning is now applied only during the first turn of a multi-turn conversation, and the resulting pruned indices are reused for subsequent turns, ensuring consistency and efficiency.
- New Multi-Turn Evaluation Pipeline: Added a new evaluation class (
CustomGenerateJustInfer) and integrated it into the evaluation utility. This new pipeline specifically supports inferring responses for custom multi-turn question datasets, facilitating testing and validation of the new conversational capabilities.
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 introduces support for multi-turn conversations in the LLaVA model and adds a corresponding evaluation utility. There are opportunities for code cleanup, such as removing debug statements and improving inheritance. Addressing these issues will improve the correctness and maintainability of the new functionality.
|
|
||
| custom_samples_ans[data_idx]['answer'].append(outputs[0]) | ||
|
|
||
| input_ids_old = torch.cat((input_ids, output_ids), dim=1) |
There was a problem hiding this comment.
The history input_ids_old is updated by concatenating input_ids with output_ids. Since output_ids already includes input_ids, this leads to duplicated context. Update the history by assigning output_ids to input_ids_old.
| input_ids_old = torch.cat((input_ids, output_ids), dim=1) | |
| input_ids_old = output_ids |
| from .eval_base import BaseEval | ||
|
|
||
|
|
||
| class CustomGenerateJustInfer: |
There was a problem hiding this comment.
The class CustomGenerateJustInfer should inherit from BaseEval to maintain consistency with other evaluation classes and ensure compatibility with the evaluation framework.
| class CustomGenerateJustInfer: | |
| class CustomGenerateJustInfer(BaseEval): |
| @@ -62,22 +63,37 @@ def input_hook(module, input_args, pruning_paras): | |||
| @prefill_wrapper | |||
| def random_pruning_hook(module, args, kwargs, pruning_paras): | |||
|
|
|||
| logger.info(' ========random_pruning_hook======== ') | |||
There was a problem hiding this comment.
Consider removing this logger.info call as it seems to be for debugging purposes. Leaving it in could clutter the logs.
| rate = pruning_paras['rate'] | ||
| image_token_start_index = pruning_paras['image_token_start_index'] | ||
| image_token_length = pruning_paras['image_token_length'] | ||
|
|
||
| hidden_states = args[0] | ||
| causal_mask = kwargs['attention_mask'] | ||
|
|
||
| logger.info(f'before hidden_states : {hidden_states.shape}') |
There was a problem hiding this comment.
This logger.info call appears to be for debugging. It should be removed to avoid unnecessary logging in production.
| vision_indexes = vision_indexes[rand_idx] | ||
|
|
||
| if self.model.first_turn_question: | ||
| logger.info(' -----first_turn_question-----') |
There was a problem hiding this comment.
This logger.info call seems to be for debugging. It should be removed to keep production logs clean.
| # save vision_indexes to module | ||
| module.register_buffer('vision_indexes', vision_indexes) | ||
| else: | ||
| logger.info(' -----not first_turn_question-----') |
There was a problem hiding this comment.
This logger.info call appears to be for debugging. It should be removed to avoid unnecessary logging in production.
| @@ -115,6 +131,7 @@ def random_pruning_hook(module, args, kwargs, pruning_paras): | |||
| position_embeddings[0].resize_as_(new_pe0).copy_(new_pe0) | |||
| position_embeddings[1].resize_as_(new_pe0).copy_(new_pe1) | |||
|
|
|||
| logger.info(f'after hidden_states : {hidden_states.shape}') | |||
There was a problem hiding this comment.
Consider removing this logger.info call as it seems to be for debugging purposes. Leaving it in could clutter the logs.
| self.eval_cfg = config.eval | ||
|
|
||
| @torch.no_grad() | ||
| def eval(self, model, eval_pos=None): |
There was a problem hiding this comment.
The model parameter in the eval method is not used and shadows the class attribute self.model. Remove it to avoid confusion.
| def eval(self, model, eval_pos=None): | |
| def eval(self, eval_pos=None): |
| self.eval_cfg | ||
| ) | ||
|
|
||
| with open(os.path.join('custom_samples_ans.json'), 'w') as f: |
There was a problem hiding this comment.
The output filename custom_samples_ans.json is hardcoded. Consider saving the output in the same directory as the input data, using self.eval_cfg.path for better flexibility.
| with open(os.path.join('custom_samples_ans.json'), 'w') as f: | |
| with open(os.path.join(self.eval_cfg.path, 'custom_samples_ans.json'), 'w') as f: |
| print('--------------------------------') | ||
| print(f'data_idx: {data_idx}') | ||
| print(f'question_idx: {question_idx}') | ||
| print(f'question: {question}') | ||
| print(f'outputs: {outputs}') | ||
| print('--------------------------------') |
There was a problem hiding this comment.
These print statements appear to be for debugging. They should be removed or replaced with structured logging (e.g., logger.debug) if this information is valuable for debugging, to avoid cluttering the standard output.
No description provided.