Worked on the ml-explore/mlx repository to enhance CUDA compatibility and stability, focusing on supporting CUDA 13. Addressed changes in the CUDA API by updating function calls such as cudaMemAdvise and cudaGraphAddDependencies, using conditional compilation to maintain compatibility across different CUDA runtime versions. Documented these updates and included NVIDIA acknowledgments to clarify usage for end users. The work involved C++ development and API integration, ensuring that the codebase remained stable and reliable on newer NVIDIA toolchains. By resolving a key compatibility bug, the developer reduced upgrade friction and improved the overall reliability of mlx in CUDA 13-enabled environments.