LAMMPS-ANI

LAMMPS-ANI interface for large scale molecular dynamics simulation with ANI neural network potential.

Quick Start

Docker

docker pull ghcr.io/roitberg-group/lammps-ani:latest
docker run --gpus all -it ghcr.io/roitberg-group/lammps-ani:latest
cd /lammps-ani/examples/water && ./run.sh

Singularity (HPC)

module load singularity
singularity pull -F docker://ghcr.io/roitberg-group/lammps-ani:latest
SINGULARITYENV_CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES singularity exec --cleanenv -H ./:/home --nv lammps-ani_latest.sif /bin/bash
cp -r /lammps-ani ./lammps-ani
cd lammps-ani/examples/water && ./run.sh

Note: The pre-built container only supports Kokkos for Ada GPUs (L4, RTX 4090). For other GPUs, see Container Guide to re-build within the container. For multi-GPU, recommend to build from source.

Examples

• water - Simple water simulation
• alanine-dipeptide - Alanine dipeptide simulation
• benchmark - Performance benchmark with water box
• combustion - Combustion reaction
• early_earth - Early Earth chemistry simulation

Set environment before running: source ./build-env.sh

Installation

Requirements

• CUDA >= 12.8.1
• GCC >= 14.2.0
• PyTorch >= 2.8.0
• OpenMPI >= 5.0.7
• CMake >= 3.21.3

Conda Environment

export conda_env_path=/path/to/env
conda create --prefix $conda_env_path python=3.11
conda activate $conda_env_path
pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128
conda install -c conda-forge libopenblas gsl -y

Build from source

Following example is for HiPerGator L4 GPU node, adjust as needed for other systems.

srun --partition=hpg-turin --cpus-per-task=30 --gres=gpu:2 --mem=200gb -t 2:00:00 --pty /bin/bash -i
export conda_env_path=/blue/roitberg/apps/torch28/
conda activate $conda_env_path
module load cuda/12.8.1 gcc/14.2.0 openmpi/5.0.7 cmake/3.21.3

git clone --recursive git@github.com:roitberg-group/lammps-ani.git
# build
cd lammps-ani && ./build.sh
# export LAMMPS_ANI_ROOT environment variable; need to run this in every new shell
source ./build-env.sh

For additional build notes, see Build Instructions.

Test

cd examples/water/
# Run a water box simulation
bash run.sh

Usage

pair_style     ani 5.1 model_file device [num_models] [ani_aev] [ani_neighbor] [ani_precision]
pair_coeff     * *

• model_file - Path to TorchANI model (.pt file)
• device - cuda or cpu
• num_models - Number of ensemble models, -1 for all (default: -1)
• ani_aev - cuaev or pyaev (default: cuaev)
• ani_neighbor - full or half (default: full)
• ani_precision - single or double (default: single)

Models

Available in /lammps-ani/models/ (container) or export with pytest models/test_models.py -s -v:

• ani2x.pt - Standard ANI-2x
• ani1x_nr.pt - ANI-1xnr Model (ani-1xnr)

For more usage details, see Usage Instructions.

Troubleshooting

B200 GPUs hang: export UCX_NET_DEVICES=mlx5_0:1

Multi-GPU in containers: Not recommended due to GPU Direct RDMA limitations. Build from source instead. See #70.