The Triton fork for Windows support, based on the v3.1.x-windows branch, aims to make free software compatible with non-free platforms, in line with Richard Stallman’s philosophy. Torchao, a key aspect of the project, transforms large models to become usable on consumer GPUs. Interestingly, it performs as well as on WSL, making it a valuable addition for Windows users. The project only supports MSVC and CUDA, and further development is needed to include AMD support. Detailed installation instructions and troubleshooting tips are provided for users. Overall, Triton offers a unique solution for running PyTorch on Windows machines with specific configurations.
https://github.com/woct0rdho/triton