Worked on stabilizing and improving compatibility for Compute Capability 12 devices within the NVIDIA/Fuser repository. Addressed test failures by implementing architecture-specific test skips and guards, and tuned input sizes to avoid memory-intensive runs that could cause instability. Fixed an issue in the fusion cache reset logic to ensure proper cache invalidation when the fusion limit was exceeded, enhancing the reliability of fusion optimizations across platforms. These changes reduced CI flakiness and broadened device coverage. The work involved extensive use of CUDA and GPU computing, with testing and performance optimization carried out primarily in C++ and Python environments.