I’m a little puzzled by the current status of the PyTorch wheels for Windows on PyPI. I’m aware the “official” way of installing PyTorch through pip on Windows is to use the index from pytorch.
I noticed the wheel for the CPU build of the main
torch package is available on PyPI, while the win wheels for
torchaudio packages are not.
I cannot find any recent helpful information about this, only some discussions in GitHub issues (mainly this and this) about binary size constraints from PyPI. Looking at the
torchaudio whls from the pytorch index, this should not be an issue, though, as they are both < 1 MB.
torchvision builds used to be pushed to PyPI (up until version 0.5.0) which makes me wonder even more why they are not anymore. On the other hand,
torchaudio has never had any Windows wheels pushed.
I understand the binary size problems related to the CUDA builds with NVIDIA not providing a global CUDA PyPI package but I do not really see a reason why the CPU builds could not be pushed to PyPI?
Does somebody have more information about whether this is intended or if it’s just a bug?