feat: Add num_workers support for DataLoader to speed up cache preparation #2165
Conversation
Summary of ChangesHello @GOavi101, 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 efficiency of the cache preparation phase by introducing support for parallel data loading via the Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Using Gemini Code AssistThe 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 by creating a comment using either
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 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
|
…ation- Add dataloader_num_workers parameter to DatasetArguments- Implement parallel data loading in format_calibration_data()- Optimize device transfers in cache preparation- Add safety checks for small datasets and custom collatorsFixes Signed-off-by: GOavi101 <[email protected]>
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces support for parallel data loading using num_workers in the DataLoader to accelerate cache preparation. The changes include adding a dataloader_num_workers argument, implementing the logic to use it in format_calibration_data with appropriate safety checks, and optimizing device transfers in IntermediatesCache. The implementation is solid, and I have one suggestion to improve code clarity.
GOavi101
force-pushed
the
fix/cache-preparation-speedup
branch
from
35f70ea to
f126d8a
Compare
Combine if/elif conditions into single if statement with or operator for better readability Signed-off-by: GOavi101 <[email protected]>
|
👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review. Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed. |
…er num_workers choice - Use args.dataloader_num_workers directly instead of getattr - Remove logic that silently ignores user-specified num_workers - Add dataloader_num_workers parameter to oneshot() function - Remove redundant device check in cache.py (torch.Tensor.to handles this) Signed-off-by: GOavi101 <[email protected]>
kylesayrs
left a comment
kylesayrs
left a comment
There was a problem hiding this comment.
Awesome, thanks for the awesome work!
| case torch.Tensor(): | ||
| return IntermediateValue( | ||
| value=value.to(device=offload_device), | ||
| value=value.to(device=offload_device) if offload_device is not None else value, |
There was a problem hiding this comment.
Is this addition necessary?
2 participants
Summary
Adds
dataloader_num_workersparameter to enable parallel data loading during cache preparation, speeding up the "Preparing cache" phase for large datasets.EDIT: Fixes #1857
Changes
dataloader_num_workersparameter toDatasetArgumentsformat_calibration_data()IntermediatesCache._offload_value()