[Q&A] Issue running MD/PIMD with PyTorch-based models in LAMMPS/i-PI #5553

Hi @dosu

1. The version 3.1.3 is already the latest, so I don't think I can upgrade it.

2. I can't use TensorFlow on the new HPC I'm on because I was not abble to install a version running on ARM Grace Hopper GPU. This is why I used pytorch.

3. I can't use : 1.1. Install DeePMD-kit in 1s to install on the HPC. Also I can't use 1.2. Install off-line packages, because it does not provide a version running on ARM.

Hi @mcourtiel, thanks for the clarification. I need to correct my earlier response — looking more carefully, v3.1.3 should already include the fix from PR #4728 (which was released in v3.0.3 back in May 2025) [1]. So if you truly have v3.1.3, the TensorFlow import error you're seeing is unexpected.

A few things to investigate:

1. Verify your installation is clean. It's possible there's a stale or conflicting installation. Try:

pip show deepmd-kit
dp --version
python -c "from deepmd.lmp import get_op_dir"

This will help confirm what's actually installed and whether the import still fails.

2. Python 3.14 may be the issue. Your traceback shows python3.14 in the paths — Python 3.14 is extremely new (still in development). It's possible that the deepmd-kit package you installed via pip is an older cached build or that some module resolution is broken on Python 3.14. I'd recommend trying with Python 3.11 or 3.12 which are well-tested with DeePMD-kit.

3. Try reinstalling in a fresh environment:

conda create -n deepmd python=3.11
conda activate deepmd
pip install torch # or install via conda for ARM
pip install deepmd-kit[lmp,ipi]

4. Build from source for ARM Grace Hopper. If pip packages don't provide proper ARM builds, you can install from source [2]:

git clone https://github.com/deepmodeling/deepmd-kit.git
cd deepmd-kit
pip install .

For the LAMMPS C++ interface with PyTorch on ARM, you'd build with CMake [3]:

cd source && mkdir build && cd build
cmake -DENABLE_PYTORCH=TRUE -DUSE_PT_PYTHON_LIBS=TRUE -DCMAKE_INSTALL_PREFIX=$deepmd_root ..
make -j4 && make install

Note: another ARM Grace Hopper user reported that LAMMPS plugin/version matching is critical — the plugin must be compiled against the exact same LAMMPS version you run [4].

Could you confirm your Python version and share the output of dp --version? The Python 3.14 in your traceback stands out as a likely root cause.

To reply, just mention @dosu.

Docs are dead. Just use Dosu.

Leave Feedback Ask Dosu about deepmd-kit Share Dosu with your team

Hi @mcourtiel,
The important point here is that the failure occurs before the .pth model is actually loaded.

Your traceback shows:

deepmd.lmp
 → deepmd.tf.env
 → import tensorflow.compat.v1

So the issue is not necessarily that your PyTorch model is incompatible with LAMMPS or i-PI. The Python entry point is importing the TensorFlow-related module during initialization, even though your model was trained and compressed with the PyTorch backend.

Since you are on an ARM Grace Hopper system where TensorFlow is not a viable option, I would avoid converting the model to .pb.

I would first verify exactly which source files are being imported:

import deepmd
import deepmd.lmp
print("deepmd:", deepmd.__file__)
print("lmp:", deepmd.lmp.__file__)

Also check the installed package and Python version:

python -m pip show deepmd-kit
python --version
python -c "import torch; print(torch.__version__)"

The traceback showing Python 3.14 is also worth investigating. If possible, I would test the same installation with a supported/stable Python version such as 3.11 or 3.12, because a very new Python version can expose package/plugin-loading issues.

For a PyTorch-only ARM environment, the most robust solution is probably to build the DeePMD-kit C++ interface and the LAMMPS plugin from source with the PyTorch backend enabled, rather than relying on the Python wheel's LAMMPS entry point.

The important things to verify in that setup are:

• PyTorch is available for the ARM architecture.
• DeePMD-kit is built with the PyTorch backend enabled.
• The LAMMPS plugin is compiled against the same LAMMPS source/version that will run the simulation.
• The resulting plugin can load the .pth model directly.

I would also test the PyTorch model independently before involving LAMMPS/i-PI, if possible, to confirm that the model itself can be loaded successfully in the installed DeePMD-kit environment.

Based on the traceback, I would focus first on the package/plugin initialization path rather than the model conversion itself.