fix sp grpo #5744
fix sp grpo #5744
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 @tastelikefeet, 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 addresses a bug in the GRPO trainer related to how it interacts with the underlying language model during padding-free operations. By dynamically identifying the correct base model, it ensures that essential forward hooks and data handling mechanisms are applied to the appropriate model component, thereby enhancing the trainer's robustness and compatibility across varied model architectures.
Highlights
- Ensure correct model reference for padding-free operations: The pull request introduces logic to dynamically determine the correct base model (either
llm_model.modelorllm_model.thinker.model) for applying padding-free hooks and methods within the GRPO trainer. - Enhance compatibility: This change improves compatibility with different LLM model structures, preventing issues when
llm_modelmight wrap the actual base model within athinkerattribute, thus making the trainer more robust.
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 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
-
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 PR addresses a bug in the grpo_trainer.py file related to handling models with a 'thinker' attribute during padding-free operations. The code changes ensure that the correct base model is used for registering forward pre-hooks and unpacking outputs, preventing potential errors when dealing with such models. The changes involve adding a conditional check for the 'thinker' attribute and assigning the appropriate base model accordingly.
| remove_handle1 = base_model.register_forward_pre_hook( | ||
| _padding_free_input_hook, with_kwargs=True, prepend=True) |
There was a problem hiding this comment.
The code now uses base_model instead of llm_model.model for registering the forward pre-hook. This change is crucial for models with the thinker attribute. It's important to verify that this change doesn't negatively impact models without the thinker attribute. Consider adding a test case to specifically cover this scenario.
| base_model._unpack_output = _padding_free_output_hook | ||
| base_model._pack_input = _padding_free_input_hook |
There was a problem hiding this comment.
Similar to the previous comment, base_model is now used for unpacking the output. Verify that this change is correct for both models with and without the thinker attribute. A test case would be beneficial here as well.
| _, inputs = base_model._pack_input(None, None, inputs) | ||
| output = base_model._unpack_output(None, None, inputs, output) |
There was a problem hiding this comment.
The code now uses base_model instead of llm_model.model for unpacking the input. This change is crucial for models with the thinker attribute. It's important to verify that this change doesn't negatively impact models without the thinker attribute. Consider adding a test case to specifically cover this scenario.
| delattr(base_model, '_unpack_output') | ||
| delattr(base_model, '_pack_input') |
There was a problem hiding this comment.
Similar to the previous comment, base_model is now used for unpacking the output. Verify that this change is correct for both models with and without the thinker attribute. A test case would be beneficial here as well.
| if hasattr(llm_model, 'thinker'): | ||
| base_model = llm_model.thinker.model | ||
| else: | ||
| base_model = llm_model.model |
There was a problem hiding this comment.
This change introduces a conditional check for the thinker attribute to determine the base_model. It's important to ensure that this logic is correct and handles all possible cases where the thinker attribute might be present or absent. Consider adding a comment explaining the purpose of this check and the expected structure of models with the thinker attribute.
| if hasattr(llm_model, 'thinker'): | ||
| base_model = llm_model.thinker.model | ||
| else: | ||
| base_model = llm_model.model |
There was a problem hiding this comment.
This change introduces a conditional check for the thinker attribute to determine the base_model. It's important to ensure that this logic is correct and handles all possible cases where the thinker attribute might be present or absent. Consider adding a comment explaining the purpose of this check and the expected structure of models with the thinker attribute.
2 participants
PR type
PR information
Write the detail information belongs to this PR.
Experiment results
Paste your experiment result here(if needed).