simulating large system with dpa finetuned model when gpu out of memory error is encountered.

[mhkcjjmt-debug]

I want to know that whether DPA (1.0 2.0 ro 3.0) models do not support compress. Running "dp --pt comperss " will encounter an error. It seems difficult to run a dpa-finetuned model on a GPU (ca. 10 G memory) when the system has 400+ atoms using lammps after the finetune process. Is there any method to simulate large system with dpa model?

[wanghan-iapcm]

for large scale md simulations, it is recommend to distill dpa-2/3 models to the compressible dpa-1 model .

please refer to https://www.nature.com/articles/s41524-024-01493-2 for the method.

please refer to https://www.bohrium.com/notebooks/76262686918 for an example.