[grpo] Support PYTORCH_CUDA_ALLOC_CONF environment variable

[hjh0119]

Support the environment variable PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True" in GRPO when using vLLM

[gemini-code-assist]

Summary of Changes

Hello @hjh0119, 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 enhances the GRPO trainer's memory management capabilities by integrating support for PyTorch's PYTORCH_CUDA_ALLOC_CONF environment variable. It introduces a mechanism to dynamically control CUDA expandable memory segments, which can help prevent out-of-memory errors and improve memory utilization, especially when working with vLLM. The changes ensure that memory allocation behavior can be fine-tuned during specific operational phases, such as vLLM engine setup and inference, while respecting user-defined environment settings.

Highlights

• CUDA Memory Management: Introduced support for the PYTORCH_CUDA_ALLOC_CONF environment variable, specifically the expandable_segments option, within the GRPO trainer when using vLLM.
• Dynamic Segment Control: Added a new utility function set_expandable_segments that allows programmatic enabling or disabling of CUDA expandable memory segments. This function respects existing user configurations of PYTORCH_CUDA_ALLOC_CONF.
• Optimized vLLM Integration: Implemented strategic calls to set_expandable_segments within grpo_trainer.py to temporarily disable expandable segments during vLLM engine initialization and inference, and re-enable them afterwards, aiming to optimize memory behavior during these critical phases.

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 by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description

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

1. 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. ↩

[gemini-code-assist]

Code Review

This pull request introduces support for toggling expandable_segments in PyTorch's CUDA allocator via the PYTORCH_CUDA_ALLOC_CONF environment variable, aiming to reduce OOM errors in GRPO training with vLLM. My review focuses on the robustness of this new state management. The current implementation manually toggles the setting, which is not exception-safe and could leave the environment in an inconsistent state. I have provided suggestions to use try...finally blocks to guarantee that the allocator settings are always restored. Additionally, I've noted a minor code redundancy in the new utility function.

[gemini-code-assist]

The global _EXPANDABLE_SEGMENTS_SET declaration is redundant. The global keyword is only necessary when assigning a new value to a global variable from within a function. Since _EXPANDABLE_SEGMENTS_SET is only being read here, this line can be safely removed to improve code clarity.