Inquiry regarding GPU acceleration for Spin-Lattice Dynamics (SLD) with MLIAPs

LAMMPS LAMMPS Development
1

Dear LAMMPS Users and Developers,

I am writing to inquire about the feasibility of running Spin-Lattice Dynamics (SLD) simulations on GPUs.

My objective is to combine SLD with Machine Learning Interatomic Potentials (MLIAPs), such as Allegro or NequIP. Since the MLIAP evaluations are significantly faster on GPU nodes compared to CPUs, I am looking to leverage GPU acceleration for the entire SLD workflow to improve simulation efficiency.

Could you please advise if the current SLD implementation (e.g., via the SPIN package) supports GPU execution, or if there are recommended workflows (such as using the Kokkos package) to achieve this integration?

Thank you for your time and assistance.

Best regards,

Zeyu

2

What is available is described in the manual. I don’t see any mention of GPU support.

Neither of these two potential implementations are part of LAMMPS, so we cannot comment on those.

3

Hi akohlmey,

Thanks for your reply. The MLIAPs can be implemented on LAMMPS either by CPU or GPU ( LAMMPS Integration - Allegro documentation ). Since currently I dont find any GPU version of the SPIN package, I find the discrepancy of calculation when combining results of atomic potentials on GPU and SPIN calculations on CPU during SLD calculations (decompositions).

Best,

Zeyu

4

As I already mentioned, those pair styles are developed and maintained outside of LAMMPS, so what they can or cannot do and how they do it is not of any concern to LAMMPS development team. We are already busy with worrying about the (currently) 92 package that are in the LAMMPS distribution.

It is what it is. Unless you want to volunteer to port the SPIN package to either the GPU package or the KOKKOS package it will likely remain like this. We don’t have the manpower (or funding) to do custom porting of packages for projects that we are not involved in. We have to rely on volunteers that need a style or package ported to GPUs and do the programming themselves and then contribute their effort. At best, people like @stamoor or @ndtrung can provide some advice on what is required. … And if you do want to volunteer to port the SPIN package, it is probably best to first contact the person that has contributed this package, @julient, and coordinate your efforts with him.

5

Dear Axel,

Thank you for the clarification and adviace.

I will look into the possibility of porting it to the GPU/KOKKOS framework.

Best regards,
Zeyu Xiang

6

Hello Zeyu Xiang,

As Axel previously mentionned, I also cannot comment on the MLIP types that are not yet ported on GPUs.

That being said, if you check within the src/KOKKOS directory, you will see that a ‘spin/kk’ atomic style is indeed defined. That means you should be able to use atoms with a ‘SPIN’ atomic style (with atom_style spin/kk) in conjunction with an ML potential that as been implemented in KOKKOS (SNAP for example).

Note that I never took the time to implement any pair/spin style (such as the exchange interation, for example…) in KOKKOS. The main reason is that evaluating those type of interactions is so much cheaper than that of the ML potential, that it did not seem very necessary.

Hope this helps,

Julien.

1 Like