Worked on the FlagTree repository to enhance GPU runtime stability and optimize memory and pipeline operations within the HCU backend. Addressed a critical issue in CUDA initialization by implementing robust error handling in C++ and Python, ensuring that runtime errors during PyTorch device retrieval no longer cause crashes but instead return safe defaults. Developed support for TLE structure primitives, including allocation, copying, and pipeline management, by updating the compiler pass pipeline and integrating LLVM conversion patterns. This work introduced advanced memory management capabilities and improved pipeline optimizations, with comprehensive integration and unit tests to ensure reliability and maintainability across backend components.