Is pytorch-cuda going to be registered on conda-forge?

Currently pytorch and torchvision are in anaconda’s conda-forge channel, but pytorch-cuda is not, so I cannot complete the install in the conda-forge channel. This is a problem under certain circumstances. Is there any plan to register pytorch-cuda in conda-forge?

I don’t believe the conda-forge copies of PyTorch, torchvision, or another PyTorch-related library are uploaded by us, but it seems by the conda-forge community/maintainers.
My guess is they might also clone new dependencies to conda-forge once the next stable release is out.

Btw. when I check it seems CPU-only binaries were cloned?
If so, then I don’t think pytorch-cuda would even be needed or are the CUDA-enabled binaries stored somewhere else on conda-forge?

Thank you @ptrblck.
I understand the situation now.
So then, in order to get the gpu version pytorch on conda-forge, I would just have to register a copy of the gpu version pytorch on the pytorch channel on the conda-forge channel.
What is the procedure to do this? Can I proceed on my own? (Can I do it?)

I don’t know how the conda-forge upload works and who is doing it, but why do you want to clone the binaries inste as d of using the already uploaded ones?

Thanks for the comment back, it appears that uploads in conda-forge are not restricted to the original library administrator or otherwise restricted. I am checking to make sure it is consistent with some pytorch-related ones that already exist.
To be more clear, I am developing a plugin for a viewer library called napari, and although the napari gui allows the library (plugin) to be installed via conda, I cannot specify any channel other than conda-forge, so I cannot install the library from the pytorch channel.

By the way, what is the actual entity of pytorch-cuda, shouldn’t it be pytorch + cudatoolkit?

pytorch-cuda is a metapackage specifying dependencies. What do you mean by “entity”?

pytorch-cuda is a metapackage specifying dependencies.

Sorry, my words were not clear, but that is exactly what I wanted to know.

1 Like