DeePMD-kit

[boegel]

• link to support ticket: #2025010260000121
• website: https://github.com/deepmodeling/deepmd-kit
• installation docs: https://docs.deepmodeling.com/projects/deepmd/en/stable/install/install-from-source.html
• toolchain: foss/2024a ?
• easyblock to use: PythonBundle
• required dependencies:
  • see https://github.com/deepmodeling/deepmd-kit/blob/r3.0/pyproject.toml
  • also LAMMPS, Horovod, MPICH, ...?
• notes:
  • ...
• effort: (TBD)
• other install methods
  • conda: yes, see https://docs.deepmodeling.com/projects/deepmd/en/stable/getting-started/install.html#install-with-conda
  • container image: yes, see https://docs.deepmodeling.com/projects/deepmd/en/stable/install/easy-install.html#install-with-docker
  • pre-built binaries (RHEL9 Linux x86_64): no
  • easyconfig outside EasyBuild: no

[boegel]

@lexming @smoors Does this ring a bell for you?

The person who requested it mentioned that it's already available on Hydra...

[boegel]

I've asked for some more input from the person who requested this: which installation they're using on Hydra @ VUB, and how they're using this exactly (since it seems like there's a bunch of options when installing it)

[boegel]

Try on top of foss/2023a with TensorFlow + PyTorch + jax as dependencies.

Try first CPU-only, then also with GPU/CUDA if that works.

Start with max. 2 hours to see how far you get.

[smoors]

@lexming @smoors Does this ring a bell for you?

The person who requested it mentioned that it's already available on Hydra...

don't think we have it, the user indicates he was wrong..

[boegel]

More details on requirements:

Considering the backends: DeePMD is most often combined with TensorFlow (tensorflow or
tensorflow-gpu), so I recommend installing it with that option. If that doesn't work,
the second option would be PyTorch although this is less preferred. Dpdata is
optional, but preferred. Compiling with C++ is a must (so it can interphase with
LAMMPS later for validation).

@pavelToman So, give it a try with TensorFlow + PyTorch + jax, and quickly focus on TensorFlow if either PyTorch or jax are problematic for some reason.