Best strategy to run mace on CPU #76
|
Hi, I am using MACE-POLAR models to run simulations using the ASE interface. I am wondering what is the correct strategy to obtain the best possible performance on CPU-only system? I am running on ARM chips with > 100 cores, so multithreading type parallelisation should be possible. Is there any way to do this via the torch? Documentation seems quite sparse and I do not exactly understand how multithreading or multiprocessing should work with MACE. I believe symmetrix has thread based parallelisation but it does not seem to be supporting models with charge/spin as of now. Would appreciate any suggestions you have. Thanks! |
Replies: 2 comments
|
Hey @shoubhikraj, at the moment there is no efficient way to run the polar model on CPU. That might arrive in a release of symmetrix if @wcwitt has the bandwith for it. |
|
@ilyes319 Thanks for the reply. Just wondering, does the pytorch compile accelerate inference with MACE-POLAR models? If so, is there any suggestion on the mode of compile to use? |
Hey @shoubhikraj, at the moment there is no efficient way to run the polar model on CPU. That might arrive in a release of symmetrix if @wcwitt has the bandwith for it.