Worked on the pytorch/FBGEMM repository to address stability concerns during a kernel migration affecting AMD hardware. Focused on maintaining consistent training throughput, the developer identified a performance regression after migrating TBE UVM cache kernels to FBGEMM_LAUNCH_KERNEL. Using C++ and CUDA, they reverted the migration to prevent production regressions, applying strong debugging and performance optimization skills. The approach included documenting the issue, outlining next steps, and planning for a corrected re-application after further testing. This work ensured that AMD deployments remained stable while a more robust solution was developed, reflecting a careful and methodical approach to risk mitigation.