kokkos suppoting issue and GRACE/FS GPU running in lammps

[Seunghyo-Noh]

Hi,

I have two questions.

[1] Currently, GRACS/FS-LAMMPS doesn't support Kokkos. Does it have update schedule to support Kokkos.
I think GRACE-FS is very efficient tool for molecular dynamics and kokkos can accelerate the simulations.

[2] GRACE/FS supports GPU in lammps, but it run in CPU when I run GRACE/FS in lammps.
How to run it using GPU. Does it need extra job or setting?

Best regards,
Seunghyo Noh

[yury-lysogorskiy]

[1] for converting foundational model to GRACE/FS YAML format for standalone C++ implementation you should use grace_utils

grace_utils -p ~/.cache/grace/checkpoint/some-grace-fs-model/model.yaml -c ~/.cache/grace/checkpoint/some-grace-fs-model/checkpoint.index export -sf

[2] Yes we do plan to include KOKKOS support for standalone GRACE/FS. However, if you have GPU, then you can use TensorFlow implementation of GRACE/FS. i.e. out-of-the-box foundational GRACE/FS models.

[3] Currently GRACE/FS standalone C++ is for CPU only, as it covers this niche when you have no GPU but many CPU cores/nodes and want to parallelize. If you have GPU, you can use TF (see p.[2]) even in combination with MPI parallelization like GRACE-1L models. See here

Thank you for your interest,
Yury

[Seunghyo-Noh]

[1] for converting foundational model to GRACE/FS YAML format for standalone C++ implementation you should use grace_utils

grace_utils -p ~/.cache/grace/checkpoint/some-grace-fs-model/model.yaml -c ~/.cache/grace/checkpoint/some-grace-fs-model/checkpoint.index export -sf

[2] Yes we do plan to include KOKKOS support for standalone GRACE/FS. However, if you have GPU, then you can use TensorFlow implementation of GRACE/FS. i.e. out-of-the-box foundational GRACE/FS models.

[3] Currently GRACE/FS standalone C++ is for CPU only, as it covers this niche when you have no GPU but many CPU cores/nodes and want to parallelize. If you have GPU, you can use TF (see p.[2]) even in combination with MPI parallelization like GRACE-1L models. See here

Thank you for your interest, Yury

@yury-lysogorskiy
Thank you for your kind reply.

I have a simple more questions about answer [2].

Do you mean the following command to use GPU including GRACE/FS
from tensorpotential.calculator import grace_fm
calc = grace_fm('GRACE-FS-OMAT')

I tried it to simulate MD by pythonic environment (ASE) and
I found that it is slower compared to GRACE/FS standalone C++ (multi-cores).
(Perhaps, it is due to ASE environment)

So, I expect it will show good performance (about speed) for large-scale simulations, if the GRACE/FS supports GPU-LAMMPS by kokkos.

Thank you again for sharing GRACE machine learning potential!

Best regards,
Seungyo Noh

[yury-lysogorskiy]

Do you mean the following command to use GPU including GRACE/FS from tensorpotential.calculator import grace_fm calc = grace_fm('GRACE-FS-OMAT')

yes

I tried it to simulate MD by pythonic environment (ASE) and I found that it is slower compared to GRACE/FS standalone C++ (multi-cores). (Perhaps, it is due to ASE environment)

mostly likely it is ASE-Python overheads