I will have to defend users here and blame AI scientist/developers together with the whole docker culture. The way they package and delivery software is just terrible. And even though pytourch has support for all GPU apis the devs seems to have a hard time adding this support. Granted, part of that is also on pytourch that for some reason Don't seem to be able to make a unified release that supports all GPUs and then you need different versions and by default IIRC they default package only supprt CUDA and CPU.