During April 2025, Xiaoli Liu developed FP8 down-cast performance optimizations and kernel stability improvements for the intel/torch-xpu-ops repository. Focusing on GPU programming and performance optimization in C++ and Python, Xiaoli enabled efficient kFloat8_e4m3fnuz and kFloat8_e5m2fnuz down-cast and up-cast copy paths, which increased FP8 throughput. The work addressed kernel reliability by resolving a build issue in the '_nocast' kernel within loop structures, reducing runtime failures. Xiaoli also strengthened unit testing to validate FP8 down-cast paths and overall copy correctness. The contributions reflect a focused, in-depth approach to performance, reliability, and maintainability within the codebase.