3 ms·
Has anyone else actually installed and demonstrated this? (might be the perfect excuse to buy an M1) "All the code the recreate the tests are here and here. Ho
by amirhirsch 5y ago
Has anyone else actually installed and demonstrated this? (might be the perfect excuse to buy an M1)
"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?
- magic_at_nodai 5y agoThis is not part of regular pytorch install. If you can build torch-mlir and SHARK from src you can use it. So hopefully soon we can make pip installable packages but for now the interfaces are in constant development so you will have to build from source. //part of nod.ai / SHARK team.
- amirhirsch 5y agoThanks for your comment, I'm betting that the M1 example demoed here should be easy to build, and I really do want the M1 for my work. I'm just lamenting how annoying it is to install the packaged versions of PyTorch for GPU since that's what I'm fighting with at this very moment!
- magic_at_enimai 5y agohear your pain and we really want to make it easy (after we make it work). //part of nod.ai / SHARK team.