PyTorch 2.14 RFC would make CUDA 13.2 the default PyPI build

PyTorch maintainers have opened an RFC to set the CUDA support matrix for release 2.14, promoting CUDA 13.2 to the default stable toolkit and the build published on PyPI.

Under the plan, pip install torch would pull CUDA 13.2. CUDA 13.0 would stay a fully CI-covered stable option via download.pytorch.org, and CUDA 12.6 would remain as a legacy channel for older GPUs. That is a shift from earlier releases, where 13.2 was experimental and 13.0 was the PyPI default.

Architecture coverage follows the toolkit lines. CUDA 12.6 still reaches back to Maxwell through Hopper. CUDA 13.0 and 13.2 drop pre-Turing hardware and add Blackwell (and extra aarch64 Blackwell variants). The stated goal is to put the newest toolkit and Blackwell improvements in front of default users after 13.2 spent 2.12 and 2.13 as experimental, while leaving a smooth path for teams not ready to move.

CUDNN pins for 13.0 and 13.2 were still marked TBD in the proposal.