In April 2026, Zach Garvey developed a HIPDNN-backed ROCm convolution backend for PyTorch in the pytorch/pytorch repository, focusing on backend development and deep learning performance. He migrated the core convolution path from MIOpen to hipDNN, introducing a more flexible kernel provider system and improving runtime efficiency. Using C++ and Python, Zach implemented a thread-local ParamsLRUCache to reduce graph build costs for repeated convolutions and expanded support for 2D and 3D convolution operations, including transposed variants. His work included end-to-end build and runtime integration, enabling faster kernel execution and a graph-based API for convolution operations on ROCm platforms.