Hi everyone — we’re planning to transition CUDA 13.0 to be the stable CUDA runtime/toolkit for the PyTorch wheels published to PyPI for the 2.11 release series.
What’s changing
- CUDA 13.0 will become the “stable” CUDA variant for PyTorch wheels published on PyPI (starting with the 2.11 release line).
- We will be publishing both x86_64 and aarch64 CUDA 13.0 binaries to PyPI.
- CUDA 12.6 and 12.8 binaries will continue to be available to download via download.pytorch.org.
Why we’re doing this
- Align the “stable” published wheels with the current CUDA/toolchain direction for the 2.11 timeframe.
- Reduce ambiguity for users choosing “the” default CUDA build on PyPI, and simplify ecosystem expectations around the recommended baseline.
Tracking / details
We’re tracking the proposal and support matrix discussion here:
Cheers,
Team PyTorch