Are there LAMMPS-SevenNet tests using acceleration packages?

[turbosonics]

Just quick question, have devs or anyone compiled LAMMPS-SevenNet with LAMMPS acceleration packages for LAMMPS-SevenNet MD simulation?

There are some LAMMPS packages as we all know:
https://docs.lammps.org/Speed_packages.html

And I wonder, if LAMMPS-SevenNet can get advantage from those packages, for GPU operation, and for CPU operation. If LAMMPS-SevenNet can get some boost with those packages with CPU or with GPU, then I think that would be really useful.

Especially, I'm curious how LAMMPS-SevenNet performs from GPU environment with GPU packages or Kokkos packages, and from CPU environment with OPENMP or Kokkos packages with openmpi...

Any comments from devs or anyone?

[YutackPark]

No I have not compiled 7net with LAMMPS acceleration packages. Currently, it won't give extra speed up as the 7net code is not aware of it.

Writing 7net + kokkos or GPU package has low priority because execution of 7net model takes > 99% of simulation time. Therefore, total speed up from acceleration packages will be negligible.

• We're currently working on accelerating 7net by itself and have some result. Please wait for the news!

[turbosonics]

Thanks, glad to hear about that.

BTW, may I ask what should I need to change if I want to compile SevenNet and LAMMPS-SevenNet? For our cluster environment, GPU nodes and CPU nodes are separated in different machine. GPU nodes have CPUs but their CPUs are not that efficient like dedicated CPU computation server.

So, if I need to compile SevenNet and LAMMPS-SevenNet to CPU cluster, what should I need to change? I'm usually compiling using venv, so I wonder is there anything I have to change when I compile both to CPU-only cluster.

[YutackPark]

Here's the related issue: #91
Currently CPU + LAMMPS has poor support. I recommend just serial code (not e3gnn_parallel.cpp/h, but only e3gnn.cpp/h). It is way more slower than you might expect as it can not leverage multi core. Compiling itself will be fine.

[turbosonics]

Well, I was able to compile and run LAMMPS SevenNet from CPU environment. But it requires a lot of CPU cores to reach just comparable level of one single GPU...

Whatever next version is, I hope you guys publish the new version as soon as possible. If that version comes with faster parallel LAMMPS MD & faster training, while maintaining the same or better accuracy for both MD & training, that would be great. Thanks!