The PyTorch development team has revealed that a preliminary version of the machine learning framework has received support for Apple Metal Performance Shaders (MPS) as a server-side computing architecture.

The developers noted that Apple Silicon-based devices feature a unified memory system that gives the GPU direct access to all memory at once. This approach makes the Mac a convenient training ground for large networks and models. The PyTorch team also claims that the update will help developers save money on cloud-based training solutions.

The innovation is available in the PyTorch Preview (Nightly) build. Installation requires macOS version 12.3 or higher. Detailed instructions are available on the PyTorch and Apple Developer website.