What is this?

This repository is a proof of concept on how we can change PyTorch's package installation for the better. It should enable a path to where we can have users specify which specific accelerator they would like to use by using the following installation command(s):

[!NOTE] This doesn't work as of today this is just showcasing what this could look like if we rolled this as the default torch package.

# Based on accelerator type
pip install torch[cpu]
pip install torch[cuda]
pip install torch[rocm]
pip install torch[xpu]

# Based on accelerator version
pip install torch[cpu-cxx11-abi]
pip install torch[cuda12-4]
pip install torch[cuda12-1]
pip install torch[rocm6-2]
pip install torch[rocm6-1]

How to build

Pre-requisites:

• uv
• python3.12 (with build module installed)

Commands to run

# Build the wheel and rename to platform specific
./generate.py

Work Todo

• Create automation to build on Linux, macOS, & Windows through Github Actions
• Automate this to automatically reflect PyTorch's binary build matrix