After we uploaded pytorch==1.7 to our channel, including the CPU versions alongside the GPU versions, it is not easy to install pytorch+cuda anymore.

For example, if I bdt create an environment for a package that depends on pytorch, and then would be willing to switch to a cuda-based version, the only way to do this is to specify the exact build number, and remove bob-devel from the environment (as that seems to introduce a pin to the CPU version). This seems to work (very long resolution):

$ bdt create -vvo <envname>  #create the environment for your package - will install cpu versions
$ conda remove -n <envname> pytorch torchvision
$ conda install -n <envname> "pytorch=1.7.0=py3.7_cuda11.0.221_cudnn8.0.3_0" "torchvision=0.8.1=py37_cu110"

This seems to work as well, but resolution is faster:

$ bdt create -vvo <envname>  #create the environment for your package - will install cpu versions
$ conda remove -n <envname> pytorch torchvision bob-devel  #remove bob-devel for faster resolution
$ conda install -n <envname> "pytorch=1.7.0=py3.7_cuda11.0.221_cudnn8.0.3_0" "torchvision=0.8.1=py37_cu110"

Everything else I tried does not seem to work.

Would it be possible to remove the CPU versions from our channel and leave only the cuda-11 versions? That would simplify deployments a lot.

@amohammadi: do you understand why this is the case?