Mati Restelli developed CUDA-aware MPI detection for Cray MPICH within the pytorch/pytorch repository, enabling GPU-direct PyTorch support on Cray supercomputers. Using C++ and leveraging both CUDA and MPI, Mati implemented a Cray-specific preprocessor branch and runtime checks for MPIX_GPU_SUPPORT_CUDA, aligning Cray MPI behavior with the existing Open MPI detection logic. This approach ensured that PyTorch could efficiently utilize GPU resources for high-performance computing workloads, reducing CPU involvement during GPU transfers. The work was validated on ALCF Polaris with NVIDIA A100s, and the codebase was extended to maintain consistent performance enhancements across different MPI implementations without introducing regressions.