Multi GPUs in lammps without Kokkos

[afour9961207]

Hi all,

I compiled the pair_allegro LAMMPS version without the Kokkos and AOTI package.

I’m wondering—can I use multiple GPUs in a single lammps run to access more gpu memory, or is that only possible with the Kokkos-enabled build of LAMMPS?

Currently, I run my simulation with the command:
lmp < input.in

Thanks in advance for your help!

Best regards,
Chang-Ti Chou

[afour9961207]

One follow-up question:

I used the new NequIP framework that was released on April 23rd, 2025, and set model_dtype: float32. Then, I followed the instructions to compile the .pth file using nequip-compile.
nequip-compile --input-path $PWD/XXX/best.ckpt --output-path ./test.nequip.pth --device cuda --mode torchscript

However, I noticed that LAMMPS still shows the following in the log file:
NequIP/Allegro is using input precision d and output precision d.
I checked the online NequIP documentation but didn’t find any guidance regarding precision or float settings.
This makes me wonder if I did something wrong. My goal is to use float32 to reduce GPU memory usage.

My version of pair_nequip_allegro is v0.7.0

Thanks for your help!

[Linux-cpp-lisp]

Good catch--you did not do anything wrong. While the model will do all of its expensive internal calculations in float32, it will still input and output float64 which ensures better numerics. This printed message is maybe confusing now, maybe we should remove it.