torch.det was the last tensor op in chgnet lacking MPS support which is why we refactored to torch.dot(lattice[0], torch.cross(lattice[1], lattice[2])) to support MPS earlier. based on tests passing on macos-14, looks like torch.det is now implemented in MPS so happy to merge.

I tested torch.det on my Apple Silicon machine with pytorch == 2.2.1. Unfortunately, it still complains thatNotImplementedError: The operator 'aten::_linalg_det.result' is not currently implemented for the MPS device.

Currently, maybe we should change the code to torch.dot(lattice[0], torch.cross(lattice[1], lattice[2], dim=-1)).

interesting, that means the tests on macos-14 aren't actually using MPS. wonder why...

edit: CI logs show CHGNet will run on cpu so indeed MPS is untested