running pytorch models without CUDA/cuDNN

[igorshmukler]

In addition to speed improvements achieved when running under TF, one obvious benefit of moving to tensorflow would be running training and inference while using tensowflow-macos and tensorflow-metal rather than torch.backends.cuda and/or torch.backends.cudnn.

Apple hardware, as well as other non nVidia devices simply will not run cudnn under torch. Therefore, currently, almost any meaningful work with torch requires nVidia hardware locally, or have a remote rig in the cloud somewhere with CUDA cards. Would be nice to be able to use a laptop for development/debugging. Unfortunately, Apple hardware has different silicon, supported by tensowflow-metal and I believe that there are no nVidia drivers for newer macOS versions.

Any ideas on the subject?

[Not to mention that some essential packages, like triton only work on Linux.]