Developed and integrated a CUDA Graphs Benchmarking Mode for the facebookresearch/param repository, enabling standardized and reproducible GPU performance analysis. The implementation introduced a new command-line argument to activate graph-based benchmarking, establishing a workflow that includes warmup, graph capture, and replay phases. Leveraging Python and CUDA, the solution ensured compatibility with both CUDA and ROCm devices while addressing challenges related to asynchronous operations for reliable measurement. This work supports cross-platform performance comparisons and accelerates optimization cycles by providing a consistent benchmarking framework. The focus on distributed systems and performance optimization contributed to more robust and efficient GPU benchmarking processes.