"All the code the recreate the tests are here and here. However you will need to install PyTorch torchvision from source since torchvision doesn’t have support for M1 yet. You will also need to build SHARK from the apple-m1-max-support branch from the SHARK repository."
So I bet you won't get it running within an hour, possibly 24 hours.
I'm testing FFTs on GPU. I got cupy running in jupyter today after a day of pain. The error said: "One method of fixing this is to repeatedly uninstall numpy until none is found" -- I had to manually delete directories and reboot to get cupy working -- now I have very fast fourier tranforms, but the internet says the PyTorch implementation is even better on GPU.... so I have to try it...
I created a new venv and tried to install pytorch with GPU support but torch.cuda.is_available() returns False.
Why is package management such a disaster?