Different package resolution with conda and mamba when installing pytorch

[thomasaarholt]

Using mamba in a fresh install of mambaforge, I went to install the pytorch CPU version today on a computer without a CUDA GPU. I was surprised to see that cudatoolkit (1GB) was included in the packages to be installed. Then by accident I did the same thing again using conda, and realised that it was excluded in that case.

The following two commands result in different output:
conda create --name torchtest pytorch torchvision torchaudio cpuonly -c pytorch
mamba create --name torchtest pytorch torchvision torchaudio cpuonly -c pytorch

To be clear: The conda result is correct. By including the cpuonly package, this somehow tells conda that the normal dependency of pytorch on cudatoolkit should be avoided. I believe it is a "mutex metapackage", like the nomkl package is. I am not familiar with how they work.

I believe this is related to #924 and may be fixed in openSUSE/libsolv#448, but I thought I'd mention it anyway.

[wolfv]

This is a duplicate. The pytorch package is not a mutex package, it's using deprecated conda "track_features". pytorch/pytorch#40213

[wolfv]

#336

[wolfv]

Would be great if you can leave a comment on the pytorch repo, or even the PR that attempts to fix this: pytorch/builder#488

[wolfv]

PS. you can use pytorch from conda-forge, it should work fine

[thomasaarholt]

Thanks, I didn't spot #336 when I searched before creating the issue!

Having a look at the pytorch builder now.