In September 2025, Vandrei developed a GPU device capability querying utility for the pytorch/pytorch repository, focusing on enhancing the CUDA module’s ability to benchmark and report NVIDIA GPU performance. Using Python and leveraging expertise in CUDA and GPU programming, Vandrei implemented new properties and methods that expose theoretical hardware limits such as FLOPs and memory bandwidth. This addition enables structured benchmarking and standardized performance analysis workflows for NVIDIA devices within PyTorch. The work addressed the need for improved hardware visibility and profiling, delivering a focused, well-integrated feature that deepens PyTorch’s support for performance optimization and device capability assessment.