NEP training not running on multiple GPU nodes

[Sonti974948]

Hi GPUMD team,

I was trying to obtain a scaling plot to compare the speed of training with increasing number of GPUs, and though I was able to obtain a good scaling within 1 GPU node (i.e. across 1,2,4 GPUs), when I submitted a job with 2 GPU nodes (i.e. 8 GPUs), the training was still using only 1 GPU node (4 GPUs).

Is there a way to train a NEP model on 2 different nodes?

Thank you so much!
Siddharth Sonti

[brucefan1983]

Thanks for the question. There is curretly only support for single-node parallelization. To extend to more nodes, MPI is needed, but we have not implemented MPI yet.

[Sonti974948]

Thank you for the clarification!