Worked on the huggingface/trl repository to enhance platform compatibility by delivering Metal Performance Shaders (MPS) support in the empty_cache function, enabling broader Apple hardware integration beyond previous device checks. Addressed an MPS-related issue in the experimental empty_cache implementation, which improved reliability and user experience for machine learning workflows on macOS devices. Collaborated closely with another contributor to ensure code quality and robust cross-team review. The work was implemented using Python and PyTorch, leveraging expertise in machine learning and software development to expand hardware support and maintain stability across the platform, while focusing on targeted improvements rather than broad feature additions.