Worked on the intel/torch-xpu-ops repository to deliver FP8 down-cast performance optimization and kernel stability improvements. Focused on refining the FP8 copy paths, the work enabled efficient down-cast and up-cast operations for kFloat8_e4m3fnuz and kFloat8_e5m2fnuz formats, increasing FP8 throughput without relying on dynamic casting. Addressed a build issue affecting the '_nocast' kernel in loop constructs, which reduced runtime failures and improved overall reliability. Enhanced validation by adding targeted unit tests to verify FP8 down-cast correctness and copy behavior. Utilized C++ and Python, applying GPU programming and performance optimization skills to strengthen maintainability and test coverage.