Question about memory scaling during training

[rees-c]

Hi,

Thanks for building these models. I noticed that the training scripts for the MP pre-trained models use small batch sizes of 16. What was the reasoning for this choice?

My application requires training on graphs with hundreds to a few thousand nodes, and I was hoping that MACE's lack of explicit triplet angle computation (as in DimeNet or GemNet) would offer more favorable memory scaling. Any insights would be greatly appreciated.

Thanks, Rees