This pair style allows you to use Allegro models from the allegro package in LAMMPS simulations. Allegro is designed to enable parallelism, and so pair_allegro supports MPI in LAMMPS. It also supports OpenMP (better performance) or Kokkos (best performance) for accelerating the pair style.
For more details on Allegro itself, background, and the LAMMPS pair style please see the allegro package and our paper:
Learning Local Equivariant Representations for Large-Scale Atomistic Dynamics
Albert Musaelian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J. Owen, Mordechai Kornbluth, Boris Kozinsky
https://www.nature.com/articles/s41467-023-36329-y
and Scaling the leading accuracy of deep equivariant models to biomolecular simulations of realistic size
Albert Musaelian, Anders Johansson, Simon Batzner, Boris Kozinsky
https://doi.org/10.1145/3581784.3627041
pair_allegro authors: Anders Johansson, Albert Musaelian.
- PyTorch or LibTorch >= 1.11.0; please note that at present we have only thoroughly tested 1.11 on NVIDIA GPUs (see #311 for NequIP) and 1.13 on AMD GPUs, but newer 2.x versions may also work. With newer versions, setting the environment variable
PYTORCH_JIT_USE_NNC_NOT_NVFUSER=1sometimes helps.
pair_style allegro
pair_coeff * * deployed.pth <Allegro type name for LAMMPS type 1> <Allegro type name for LAMMPS type 2> ...
where deployed.pth is the filename of your trained, deployed model.
The names after the model path deployed.pth indicate, in order, the names of the Allegro model's atom types to use for LAMMPS atom types 1, 2, and so on. The number of names given must be equal to the number of atom types in the LAMMPS configuration (not the Allegro model!).
The given names must be consistent with the names specified in the Allegro training YAML in chemical_symbol_to_type or type_names. Typically, this will be the chemical symbol for each LAMMPS type.
To run with Kokkos, please see the LAMMPS Kokkos documentation. Example:
mpirun -np 8 lmp -sf kk -k on g 4 -pk kokkos newton on neigh full -in in.scriptto run on 2 nodes with 4 GPUs each.
We provide an experimental "compute" that allows you to extract custom quantities from Allegro models, such as polarization. You can extract either global or per-atom properties with syntax along the lines of
compute polarization all allegro polarization 3
compute polarizability all allegro polarizability 9
compute borncharges all allegro/atom born_charge 9 1
The name after allegro[/atom] is attempted extracted from the dictionary that the Allegro model returns. The following number is the number of elements after flattening the output. In the examples above, polarization is a 3-element global vector, while polarizability and Born charges are global and per-atom 3x3 matrices, respectively.
For per-atom quantities, the second number is a 1/0 flag indicating whether the properties should be reverse-communicated "Newton-style" like forces, which will depend on your property and the specifics of your implementation.
Note: For extracting multiple quantities, simply use multiple commands. The properties will be extracted from the same dictionary, without any recomputation.
Note: The group flag should generally be all.
Note: Global quantities are assumed extensive and summed across MPI ranks. Keep ghost atoms in mind when trying to think of whether this works for your property; for example, it does not work for Allegro's global energy if there are non-zero energy shifts, as these are also applied to ghost atoms.
git clone --depth=1 https://github.com/lammps/lammpsor your preferred method.
(--depth=1 prevents the entire history of the LAMMPS repository from being downloaded.)
git clone https://github.com/mir-group/pair_allegroor by downloading a ZIP of the source.
From the pair_allegro directory, run:
./patch_lammps.sh /path/to/lammps/If you have PyTorch installed and are NOT using Kokkos:
cd lammps
mkdir build
cd build
cmake ../cmake -DCMAKE_PREFIX_PATH=`python -c 'import torch;print(torch.utils.cmake_prefix_path)'`If you don't have PyTorch installed OR are using Kokkos, you need to download LibTorch from the PyTorch download page. Ensure you download the cxx11 ABI version if using Kokkos. Unzip the downloaded file, then configure LAMMPS:
cd lammps
mkdir build
cd build
cmake ../cmake -DCMAKE_PREFIX_PATH=/path/to/libtorchCMake will look for MKL automatically. If it cannot find it (MKL_INCLUDE_DIR is not found) and you are using a Python environment, a simple solution is to run conda install mkl-include or pip install mkl-include and append:
-DMKL_INCLUDE_DIR="$CONDA_PREFIX/include"
to the cmake command if using a conda environment, or
-DMKL_INCLUDE_DIR=`python -c "import sysconfig;from pathlib import Path;print(Path(sysconfig.get_paths()[\"include\"]).parent)"`
if using plain Python and pip.
CMake will look for CUDA and cuDNN. You may have to explicitly provide the path for your CUDA installation (e.g. -DCUDA_TOOLKIT_ROOT_DIR=/usr/lib/cuda/).
Note that the CUDA that comes with PyTorch when installed with conda (the cudatoolkit package) is usually insufficient (see here, for example) and you may have to install full CUDA seperately. A minor version mismatch between the available full CUDA version and the version of cudatoolkit is usually not a problem, as long as the system CUDA is equal or newer. (For example, PyTorch's requested cudatoolkit==11.3 with a system CUDA of 11.4 works, but a system CUDA 11.1 will likely fail.) cuDNN is also required by PyTorch.
pair_allegro supports the use of OpenMP to accelerate certain parts of the pair style, by setting OMP_NUM_THREADS and using the LAMMPS OpenMP package.
pair_allegro supports the use of Kokkos to accelerate the pair style on the GPU and avoid host-GPU transfers.
pair_allegro supports two setups for Kokkos: pair_style and model both on CPU, or both on GPU. Please ensure you build LAMMPS with the appropriate Kokkos backends enabled for your usecase. For example, to use CUDA GPUs, add:
-DPKG_KOKKOS=ON -DKokkos_ENABLE_CUDA=ON
to your cmake command. See the LAMMPS documentation for more build options and how to correctly run LAMMPS with Kokkos.
make -j$(nproc)This gives lammps/build/lmp, which can be run as usual with /path/to/lmp -in in.script. If you specify -DCMAKE_INSTALL_PREFIX=/somewhere/in/$PATH (the default is $HOME/.local), you can do make install and just run lmp -in in.script.
-
Q: My simulation is immediately or bizzarely unstable
A: Please ensure that your mapping from LAMMPS atom types to NequIP atom types, specified in the
pair_coeffline, is correct, and that the units are consistent between your training data and your LAMMPS simulation. -
Q: I get the following error:
instance of 'c10::Error' what(): PytorchStreamReader failed locating file constants.pkl: file not foundA: Make sure you remembered to deploy (compile) your model using
nequip-deploy, and that the path to the model given withpair_coeffpoints to a deployed model.pthfile, not a file containing only weights likebest_model.pth. -
Q: I get the following error:
instance of 'c10::Error' what(): isTuple()INTERNAL ASSERT FAILEDA: We've seen this error occur when you try to load a TorchScript model deployed from PyTorch>1.11 in LAMMPS built against 1.11. Try redeploying your model (retraining not necessary) in a PyTorch 1.11 install.
-
Q: I get the following error:
Exception: Argument passed to at() was not in the mapA: We now require models to have been trained with stress support, which is achieved by replacing
ForceOutputwithStressForceOutputin the training configuration. Note that you do not need to train on stress (though it may improve your potential, assuming your stress data is correct and converged). If you desperately wish to keep using a model without stress output, there are two options: 1) Remove lines that look like these in your version ofpair_allegro[_kokkos].cpp2) Redeploy the model with an updated config file, as described here.
