How to run calculations in parallel on GPU with a trained mace potential

[nihil39]

Hi,

as per the title, I have a MACE trained model and I don't know how to parallelize the calculation. These are the relevant line in my script:

calculator = MACECalculator(model_paths = f'{model_path}/{model_name}', device = 'cuda', dispersion = True, enable_cueq = True)
init_conf.calc = calculator

I am running these calculation on a cluster with Gpus, slurm is used as a queue manager.

As far as I understood ASE will try to run automatically in parallel if it founds mpi4py libraries in the system.

Does MACE calculators support parallel running on gpu?

Is there any environment variable that has to be exported in order to run in parallel using gpus? Like OMP_NUM_THREADS for cpu runs?

Thanks for any help.

[ilyes319]

What kind of parallelisation do you want to do? Do you want to run multiple computations in parallel, or parallels a single computation across multiple GPUs?

The first thing is doable, but the second is not yet available publicly.

[zeayng]

Is single computation parallel processing on multiple GPUs now achievable?

[gabor1]

Both Symmetrix (the package by @wcwitt ) and MLIAP will let you use multiple GPUs for a single computation with LAMMPS

[rajorshichat]

Is parallel processing of single computation across multiple GPUs also available in ASE ?

[nihil39]

Thanks, my idea was to do the second type of computation like the kind described here.
I was thinking about doing strong scaling test of the system. I also tested the cpu parallelism but, on this cluster, even 112 CPU cores are 20 times slower than 1 GPU.
I do the first type of calculation, maybe naively, just by running multiple job scripts in parallel on different gpu nodes.

So the single parallel calculation is brewing, looking forward to its release.

[ilyes319]

If you wait a few weeks, @wcwitt will release exactly what you need, stay tuned.