Build variants

A kernel can be compliant for a specific compute framework (e.g. CUDA) or architecture (e.g. x86_64). For compliance with a compute framework and architecture combination, all the build variants listed below must be available. This list will be updated as new PyTorch versions are released.

CPU aarch64-darwin

Metal aarch64-darwin

CPU aarch64-linux

CUDA aarch64-linux

CPU x86_64-linux

CUDA x86_64-linux

ROCm x86_64-linux

XPU x86_64-linux

Python-only kernels

Kernels that are in pure Python (e.g. Triton kernels) only need to provide one or more of the following variants:

Update on GitHub