Questions on the training with cueq flag enabled

[nihil39]

Hi,

I am experimenting with the cueq flag. This is the log file of the training with the cueq flag enabled. My training completes just fine and I see a speedup of more than 2x compared with the same training with the cueq flag not enabled.

Two questions:

1. What is the meaning of the following line

2025-01-08 11:20:31.490 INFO: CUDA implementation not available: Only descriptors with 3 or 4 operands are supported. Got u,u,u,u,u. ?

2. This is the log file of the same training performed without the cueq flag. Is it normal that the loss values are different compared to the ones that I see in the other training with the cueq flag enabled? Is this caused by a difference in numerical precision when using cuequivariance libraries?

Thanks for any help

[ilyes319]

Hey @nihil39 ,that warning should be fine. For the small differences, I am not sure, are you training in float32?