lmp in 2.1.3 make unrecognized pair error

[pee8379]

Bug summary

I try to continue training model (from deepmd-kit 2.1.1, dpgen 0.10.5), using newly released 2.1.3. (offline package deepmd-kit-2.1.3-cuda10.2_gpu-Linux-x86_64.sh )

dpgen process failed to run, (task 4) with [unrecognized pair style] error in lmp log file.

DeePMD-kit Version

2.1.3

TensorFlow Version

How did you download the software?

Offline packages

Input Files, Running Commands, Error Log, etc.

dpgen 0.10.5
call lmp by
"model_devi":[ "command": "/DeepMD-kit-location/deepmd-kit-2.1.1-cu10.2/bin/lmp"
in machine json file.

Steps to Reproduce

Further Information, Files, and Links

No response

[njzjz]

Please activate the environment first:

conda activate /path/to/deepmd-kit

（source activate /path/to/deepmd-kit or source /path/to/deepmd-kit/bin/activate /path/to/deepmd-kit may also work)

By activating the conda environment, several environment variables can be set correctly.

[r-hou]

it helps nothing, I did activate the environment, otherwise how could I use "lmp -h", it doesn't show deepmd pair_style

[njzjz]

it helps nothing, I did activate the environment, otherwise how could I use "lmp -h", it doesn't show deepmd pair_style

Currently, lmp -h only lists built-in styles, but no plugged styles. This part of the codes is maintained by LAMMPS.

[hyuqing11]

I have activated the environment but I still get the error message: "ERROR: Unrecognized pair style 'deepmd' (src/force.cpp:271)" Do you have any suggestions?

[njzjz]

I have activated the environment but I still get the error message: "ERROR: Unrecognized pair style 'deepmd' (src/force.cpp:271)" Do you have any suggestions?

@hyuqing11 If it was installed from conda-forge or the offline package released recently, the problem has been resolved in conda-forge/deepmd-kit-feedstock#63.

---

During the activation process, the environment variable LAMMPS_PLUGIN_PATH will be set, so it can also be manually set instead:

export LAMMPS_PLUGIN_PATH=$CONDA_PREFIX/lib/deepmd_lmp

[hyuqing11]

Thank you for your help. However, I encountered an issue after setting the LAMMPS_PLUGIN_PATH. The error message I received is as follows: "Open of file /test_env4/lib/deepmd_lmp/dpplugin.so failed: libabsl_status.so.2301.0.0: cannot open shared object file: No such file or directory." Do you have any suggestions?

[chazeon]

@njzjz @hyuqing11

I have encounter the same kind of problem today.
The deepmd_lmp/dpplugin.so from the current conda-forge version of deepmd-kit does not has proper rpath and needed shared library information.

After installation, a temporary patch can be done with patchelf by

cd $CONDA_PREFIX/lib
patchelf --add-rpath '$ORIGIN/..' deepmd_lmp/dpplugin.so
patchelf --add-needed liblammps.so deepmd_lmp/dpplugin.so

My deepmd-kit version and build is

deepmd-kit                2.2.7           cuda118py310h49a02c8_3    conda-forge

[njzjz]

The deepmd_lmp/dpplugin.so from the current conda-forge version of deepmd-kit does not has proper rpath and needed shared library information.

I know @Yi-FanLi had this issue before, but he compiled LAMMPS himself. @chazeon Did you use conda-forge's LAMMPS or others?

[chazeon]

I know @Yi-FanLi had this issue before, but he compiled LAMMPS himself. @chazeon Did you use conda-forge's LAMMPS or others?

Yes, I used conda-forge’s LAMMPS.

[njzjz]

Yes, I used conda-forge’s LAMMPS.

Do you use lmp_mpi or lmp_serial? lmp_serial in conda-forge does not provide a shared library (liblammps.so is built for MPI), which might be a problem and should not be used.

[Yi-FanLi]

Hi @chazeon, I used the following command to set up an environment:

conda create -n dp2.2.7-gpu -c conda-forge deepmd-kit=*=cuda* lammps python=3.9 pip

and used this command to run the task:

module purge
module load anaconda3/2020.11
conda activate dp2.2.7-gpu
lmp_mpi -in in.lammps

It worked fine.
Can you tell us how we can reproduce the issue?

[chazeon]

@njzjz @Yi-FanLi
I see. Currently, lmp_mpi indeed loads liblammps.so automatically and then loads the plugin correctly. I was fixing it because ldd or readelf does not list liblammps.so as a dependency.

It's my fault. I spotted the missing .so because an earlier incompatible version of LAMMPS was installed by conda because of other dependencies pinned down the LAMMPS version. I tried to debug the .so file to look into its symbols. I wasn't aware that upgrading LAMMPS version has already fixed these issues.