PyTorch is a well-known library in the world of machine learning. Open source, it is based on Torch, a library published by Meta. Unlike the latter, PyTorch operates in Python.
The development team presented last night a preview version of the library optimized for Apple Silicon, more precisely designed to exploit the GPU part of these chips via Metal, with the help of Apple.
Although this is only a preview, the results are to match. On an M1 Ultra, performance is multiplied by 5 to 20 depending on the tasks.
All developers can use this version, provided they have at least macOS 12.3, a Mac M1 (and its variants) and the arm64 version of Python.