Cen Zhao developed a CUDA Graphs Benchmarking Mode for the facebookresearch/param repository, introducing a standardized workflow for GPU performance analysis. Using Python and leveraging CUDA and ROCm technologies, Cen implemented a new argument to enable graph-based benchmarking, which captures and replays CUDA operations through a defined sequence of warmup, graph capture, and replay steps. The approach addressed asynchronous operation handling to ensure reliable measurement and device compatibility. This work established a reproducible benchmarking process, facilitating cross-hardware performance comparisons and supporting distributed systems research. Cen’s contribution provided a foundation for faster optimization cycles and more consistent performance evaluation in GPU environments.