Worked on the pytorch/pytorch repository to deliver initial ROCm support for the new gfx1250 (CDNA5) architecture, enabling early workloads and laying the foundation for future AMD GPU compatibility. The approach involved expanding GEMM and memory paths, raising shared-memory limits, and refactoring architecture-check logic for maintainability across CUDA and CUBLAS modules. Leveraged CMake, C++, and GPU programming expertise to isolate CK GEMM in a separate library and integrate AOTriton-based flash attention for gfx1250. Documented CI constraints and aligned test gating, ensuring robust validation. This work established multi-architecture build readiness and improved the pathway for performance enhancements on next-generation hardware.