why do we use miniconda with pytorch cuda image?

[haiderasad]

hi i am kind of stuck with docker gpu pytorch, basically my image is successfully built but on inference invocation when using cuda is gives out the following error
ImportError: /root/.cache/torch_extensions/py38_cu113/fused/fused.so: cannot open shared object file: No such file or directory

i am using nvidia/ image

also i would like to ask why do we use miniconda with pytorch cuda image?

[anibali]

We use Conda because it makes it much easier to install third party packages, especially when they include binaries. The PyTorch developers seem to agree, as can be seen on the PyTorch website:

Anaconda is our recommended package manager

[haiderasad]

Ok that explains a lot as i was also having troubles with loading cuda binaries from the runtime version of nvidia docker pytorch then it was solved by using the devel version , thanks