Worked on the ROCm/FBGEMM repository to enhance performance benchmarking for inference kernels by introducing a warm-up method and integrating the Kineto profiler. This approach stabilized timing measurements and enabled more accurate profiling of kernel execution time and bandwidth, reducing measurement overhead and improving the reliability of benchmarking results. Leveraging C++, Python, and GPU computing expertise, the developer focused on performance optimization and profiling to support more informed tuning and optimization decisions. The work provided a robust foundation for precise performance analysis, allowing future development efforts to better target bottlenecks and improve inference efficiency within the ROCm/FBGEMM codebase.