Installating DeePMD and DP-GEN in the same Conda environment

[RemiLacroix-IDRIS]

Hello,

I am trying to install both DeePMD and DP-GEN in the same Conda environment but so far I have been struggling to make that work.

I tried to install from the official DeepModeling Conda channel but it seems the DP-GEN version available there is pretty old and I'm getting some conflicts.

I thought I would use the Conda-Forge channel but I'm not sure that the DeePMP build available there supports multi-GPU multi-nodes configuration as Horovod is not available.

Do you have any suggestions?

Best regards,
Rémi Lacroix

[njzjz]

If you don't mind mixing conda and pip, you can use pip directly.

as Horovod is not available.

It's reasonable to have Horovod in the conda-forge.

[RemiLacroix-IDRIS]

Ok.

Now we would like to be able to use the DeePMD plugin for Lammps but the version from "conda-forge" channel does not seem to be patched. Is the version from the "deepmodeling" channel patched?

What would you recommend in our specific case where we want to have DeePMD and DP-GEN installed in the same environment while also providing a patched version of LAMMPS?

[njzjz]

Both channels do not patch LAMMPS. The plugin mechanism is used.

Several days ago I merged conda-forge/deepmd-kit-feedstock#63 to fix the
problem that the LAMMPS plugin environment variable was not set automatically. You may check if your version is no older than this fix.

[RemiLacroix-IDRIS]

Was the conda-forge channel updated already? My version does not seems to have the fix but when trying to update with conda update -c conda-forge lammps, it does not find any new versions bit it tries to downgrade my TensorFlow packages to CPU...

Here is the version currently installed:

lammps 2023.08.02 cpu_py311_hf0e8f25_openmpi_7 conda-forge

[njzjz]

deepmd-kit was updated, not lammps

[RemiLacroix-IDRIS]

Sorry, I misunderstood, I can confirm this issue is fixed. So appart from the TensorFlow workaround, everything works fine I think.