Is it possible to run single precision MD with easy-installed deepmd and LAMMPS?

[cihe13375]

Hi,

I'm trying to run a MD simulation on LAMMPS with deepmd models and I'm getting errors like this:
2022-09-11 15:07:48.220590: F tensorflow/core/framework/tensor.cc:725] Check failed: dtype() == expected_dtype (1 vs. 2) double expected, got float

After searching in discussions it seems that this error raises because the model is in single precision (yes), but LAMMPS is built with double precision. I guess this can be solved by manually compile dpmd and LAMMPS (as illustrated in the doc), but I'm wondering if it's possible to run single precision simulations without manually compiling (i.e. by offline package or conda). It seems that the conda repo doesn't have _low variants anymore starting from version 2.0 (I searched by conda search '*deepmd*[channel=https://conda.deepmodeling.com]'). Is this intentional (like, maybe single precision is discouraged for general purposes)?

I have export DP_INTERFACE_PREC=low set before both training and MD.

Thanks!

[njzjz]

In the latest version, one can change the environment variable LAMMPS_PLUGIN_PATH from something like /opt/deepmd-kit/lib/deepmd_lmp to /opt/deepmd-kit/lib/deepmd_lmp_low to switch to single precision.

[cihe13375]

Thanks, it works.