🐛 [BUG] Problem with nequip_calculator in ASE about nequip HOT 4 CLOSED

Hi @tbraeckevelt ,

Thanks for your interest in our code!

This is a known issue upstream in PyTorch-- see discussion on another NequIP issue here.

Long story short: upgrading to PyTorch 1.10 where the underlying issue is fixed should resolve this. (Also you don't need torchvision or torchaudio unless you're using them for something else.)

from nequip.

(Unrelated: I noticed in your config that you have set allow_tf32: true--- please note that we do not at this point recommend this as we have not yet fully understood the impact of reduced TF32 precision on the model or, in particular, its equivariance. If you have a reason to set this it should be fine, but if it somehow got set by accident you probably want it set to False.)

from nequip.

Ah my bad, I should have checked the closed issues. It works now, thanks!

Thanks for the tip, I have set allow_tf32 to false.

from nequip.

@tbraeckevelt great, glad to hear. And no worries this is a pretty odd error message to look up.

Closing this, feel free to open a new issue/discussion if you have other questions!

from nequip.