LAMMPS
LAMMPS stands for Large-scale Atomic/Molecular Massively Parallel Simulator and is copywrited by Sandia (DOE). See the On-line Manual for LAMMPS documentation, which is linked below. Here we give instructions how to run it on the CPUs and GPUs with the staff supported version on the Hazel HPC. Also given are instructions for users to install their own version.External Links:
LAMMPS website
LAMMPS Tutorials
LAMMPS benchmarks
Typically, it's not a good idea to run lammps from your /home directory, as its output can easily overflow your quota.
The current staff-supported version is 22Jul2025 (stable_22Jul2025_update4), compiled with PrgEnv-intel/2025.3-slurm and cuda/12.0. The binary is named lmp_hazel, installed at /usr/local/apps/lammps/22Jul2025/intel/gpu/bin/lmp_hazel.
The modulefiles are:
module avail lammps/22Jul2025
They are:
lammps/22Jul2025/intel/gpu-slurm (for GPU nodes)
lammps/22Jul2025/intel/cpu-slurm (for CPU nodes)
Previous version (2024Aug29):
CPU only, no GPU support. This version predates the Slurm-integrated compiler stack.
Older versions (2022Jun23):
These are the previous GPU-capable staff-supported builds. Prior to the 22Jul2025 build, these were the only staff-supported versions with GPU acceleration.
Modulefiles:
- lammps/2022Jun23/intel/gpu (recommended — Intel compiler, GPU nodes)
- lammps/2022Jun23/intel/cpu (Intel compiler, CPU nodes)
- lammps/2022Jun23/gnu/gpu (GNU compiler, GPU nodes)
- lammps/2022Jun23/gnu/cpu (GNU compiler, CPU nodes)
Earlier versions (29Sep21u3, 7Aug19, 22Aug2018) are no longer supported.
Table of Contents - Select for a Particular Topic
To test LAMMPS on the 22Jul2025 version for GPUs (interactively) do
salloc --ntasks=16 --nodes=1 --time=00:05:00 -p gpu_partners --qos=short_gpu --gres=gpu:a10:1 module load lammps/22Jul2025/intel/gpu-slurm srun lmp_hazel -sf gpu -in \ /usr/local/apps/lammps/22Jul2025/intel/gpu/share/lammps/bench/in.lj
To test LAMMPS 22Jul2025 for GPUs in batch mode, create "lammps_gpu_benchmark_slurm.sh" with the following lines:
#!/bin/bash #SBATCH --ntasks=16 #SBATCH --nodes=1 #SBATCH --time=00:05:00 #SBATCH --partition=gpu_partners #SBATCH --qos=short_gpu #SBATCH --gres=gpu:a10:1 #SBATCH --output=out.%j #SBATCH --error=err.%j module load lammps/22Jul2025/intel/gpu-slurm srun lmp_hazel -sf gpu -in /usr/local/apps/lammps/22Jul2025/intel/gpu/share/lammps/bench/in.ljand run it with:
sbatch lammps_gpu_benchmark_slurm.sh
Available GPU types can be checked with:
si --gpus --qos short_gpu # Note, GPUs in other QOS (e.g., the "gpu" QOS) can be checked too, but for short jobs (less than 2 hours) it is generally advisable to use the high priority short_gpu QOS.The above command shows the GPU type and availability. Example output:
Partition GPU GrpTRES Used Avail Alloc Total gpu_partners A10 - - 8 0 8 gpu_partners L40S 20 0 7 17 24 gpu_partners H200 4 0 4 4 8 gpu_partners RTX_6000_PRO - - 11 13 24Use the GPU type from the second column in your --gres flag. For example, to select the A10 GPU, use --gres=gpu:a10:1
salloc --ntasks=16 --nodes=1 --time=00:10:00 \ --partition=compute_partners --qos=short --constraint=cascadelake module load lammps/22Jul2025/intel/cpu-slurm srun lmp_hazel -in \ /usr/local/apps/lammps/22Jul2025/intel/gpu/share/lammps/bench/in.ljTo do the same thing in batch mode, create "parallel_cpu_slurm.sh" with the following lines:
#!/bin/bash #SBATCH --ntasks=16 #SBATCH --nodes=1 #SBATCH --time=00:10:00 #SBATCH --partition=compute_partners #SBATCH --qos=short #SBATCH --constraint=cascadelake #SBATCH --output=out.%j #SBATCH --error=err.%j module load lammps/22Jul2025/intel/cpu-slurm srun lmp_hazel -in /usr/local/apps/lammps/22Jul2025/intel/gpu/share/lammps/bench/in.ljRun this batch job with:
sbatch parallel_cpu_slurm.sh
Here is the link to the Hazel benchmarks.
lj model, cascadelake CPU node, 1 MPI rank (serial):
Performance: 1809.691 tau/day, 4.189 timesteps/s
lj model, cascadelake CPU node, 16 MPI ranks (parallel):
Performance: 26643.269 tau/day, 61.674 timesteps/s
lj model, A30 2x GPU node, 16 MPI ranks:
Performance: 141252.926 tau/day, 326.974 timesteps/s
Users are encouraged to install and maintain their own version of LAMMPS in:
/usr/local/usrapps/projectidWhere "projectid" is a directory name given to the project by the HPC administrator (a request must be made). The following instructions describe how to compile LAMMPS from source using the stable branch. You can choose which packages to include at the CMake configuration step.
Step 1 — Get the source
git clone -b stable https://github.com/lammps/lammps.git src cd src && git describe --tags
Step 2 — Load the build environment
module purge module load PrgEnv-intel/2025.3-slurm module load cuda/12.0 which mpiicx mpiicpx mpiifx nvcc g++ # all must resolve before continuing
Step 3 — Configure with CMake
Choose your packages with-D PKG_<NAME>=yes:
mkdir build && cd build cmake \ -D CMAKE_INSTALL_PREFIX=/usr/local/usrapps/[your_path]/lammps \ -D CMAKE_BUILD_TYPE=Release \ -D BUILD_MPI=yes -D BUILD_OMP=yes \ -D CMAKE_C_COMPILER=$(which mpiicx) \ -D CMAKE_CXX_COMPILER=$(which mpiicpx) \ -D CMAKE_Fortran_COMPILER=$(which mpiifx) \ -D CMAKE_CUDA_HOST_COMPILER=$(which g++) \ -D CMAKE_CXX_FLAGS="-march=native" \ -D CMAKE_C_FLAGS="-march=native" \ -D LAMMPS_MACHINE=[your_machine_name] \ -D FFT=MKL \ -D PKG_GPU=yes -D GPU_API=cuda \ -D GPU_ARCH="sm_61;sm_75;sm_80;sm_86;sm_89;sm_90" \ -D GPU_PREC=mixed \ -D PKG_INTEL=yes \ -D PKG_OPENMP=yes \ -D PKG_KSPACE=yes \ -D PKG_MOLECULE=yes \ -D PKG_MANYBODY=yes \ -D PKG_RIGID=yes \ # add or remove -D PKG_<NAME>=yes lines for any other packages you need # full list of available packages: https://docs.lammps.org/Packages_list.html ../cmakeAfter cmake runs, check the summary output to confirm your packages are listed under "Enabled packages" before building.
Step 4 — Build and install
make -j 8 make install
Note: Build in /tmp rather than your install directory to avoid hitting inode quotas on /usr/local. Delete the build tree after installation with rm -rf /tmp/your-build-dir.
The following alternative installation methods (conda and makefile) are preserved for reference and will be re-validated in a future update.
As of 22Jul2025 (stable_22Jul2025_update4), compiled with PrgEnv-intel/2025.3-slurm and cuda/12.0:
GPU, INTEL, KSPACE, MANYBODY, MOLECULE, OPENMP, RIGID
NequIP is a neural network interatomic potential that can be used as a LAMMPS pair style. It is built on PyTorch and runs on the GPU nodes. The workflow below installs PyTorch and NequIP with pip in a conda environment, builds LAMMPS with the NequIP pair style, trains a model, and runs an MD simulation with it.
Step 1 — Create the conda environment
Create "pyenv.yml" in /usr/local/usrapps/$GROUP/$USER with the following lines:name: test0 dependencies: - python=3.12 - pipThen create the environment and install PyTorch and NequIP with pip:
conda env create -f pyenv.yml -p /usr/local/usrapps/$GROUP/$USER/python_base conda activate /usr/local/usrapps/$GROUP/$USER/python_base python -m pip install torch==2.6.0 torchvision python -m pip install nequip python -m pip install wandb
Step 2 — Build LAMMPS with the NequIP pair style
Load required modules and build LAMMPS in /usr/local/usrapps/$GROUP/$USER/build_lammps:module load cuda/12.0 module load PrgEnv-intel/2025.3-slurm module load cmake/4.1.1 module load mkl/2024.1 conda activate /usr/local/usrapps/$GROUP/$USER/python_base mkdir /usr/local/usrapps/$GROUP/$USER/build_lammps cd /usr/local/usrapps/$GROUP/$USER/build_lammps git clone -b stable https://github.com/lammps/lammps.git git clone https://github.com/mir-group/pair_nequip_allegro.gitPatch LAMMPS with the NequIP pair style:
cd pair_nequip_allegro ./patch_lammps.sh ../lammps/ cd ..Configure and build:
cd lammps mkdir build && cd build cmake \ -D CMAKE_PREFIX_PATH=$(python -c 'import torch; print(torch.utils.cmake_prefix_path)') \ -D CMAKE_BUILD_TYPE=Release \ -D CMAKE_CXX_STANDARD=17 \ -D BUILD_MPI=yes -D BUILD_OMP=yes \ -D CMAKE_C_COMPILER=$(which mpiicx) \ -D CMAKE_CXX_COMPILER=$(which mpiicpx) \ -D CMAKE_Fortran_COMPILER=$(which mpiifx) \ -D CMAKE_CXX_FLAGS="-march=native" \ -D CMAKE_C_FLAGS="-march=native" \ -D FFT=MKL \ -D PKG_INTEL=yes \ -D PKG_OPENMP=yes \ -D PKG_KSPACE=yes \ -D PKG_MOLECULE=yes \ -D PKG_MANYBODY=yes \ -D PKG_RIGID=yes \ ../cmake make -j 8
Step 3 — Train the NequIP model
Set up a working directory on /share and get the NequIP example configs:mkdir -p /share/$GROUP/$USER/lammps cd /share/$GROUP/$USER/lammps git clone https://github.com/mir-group/nequip.git nequip_sourceDownload the training data on the login node — compute nodes have no internet access:
conda activate /usr/local/usrapps/$GROUP/$USER/python_base cd /share/$GROUP/$USER/lammps/nequip_source/configs wget "https://archive.materialscloud.org/records/ycbvx-knj69/files/fcu.xyz?download=1" -O fcu.xyzNote: the MaterialsCloud server can be slow and this download may take 5-10 minutes.
Check which GPUs are idle, then set --gres in the batch script to match. Recommended GPU for this task is H200.
si --gpus --qos short_gpu # Note, GPUs in other QOS (e.g., the "gpu" QOS) can be checked too, but for short jobs (less than 2 hours) it is generally advisable to use the high priority short_gpu QOS.Create "submit.sh" in /share/$GROUP/$USER/lammps with the following lines:
#!/bin/bash #SBATCH --job-name=nequip_train #SBATCH --partition=gpu_partners #SBATCH --gres=gpu:h200:1 #SBATCH --qos=short_gpu #SBATCH --ntasks=1 #SBATCH --cpus-per-task=4 #SBATCH --time=02:00:00 source ~/.bashrc module load cuda/12.0 conda activate /usr/local/usrapps/$GROUP/$USER/python_base export WANDB_MODE=offline cd /share/$GROUP/$USER/lammps/nequip_source/configs nequip-train -cn tutorial.yamlRun this batch job with:
sbatch submit.sh
Step 4 — Compile the trained model
Locate the checkpoint file in the outputs directory and compile it by running this on the login node after training finishes:conda activate /usr/local/usrapps/$GROUP/$USER/python_base cd /share/$GROUP/$USER/lammps/nequip_source/configs nequip-compile \ outputs/<date>/<time>/best.ckpt \ compiled_model.nequip.pth \ --device cpu \ --mode torchscript
Step 5 — Prepare the LAMMPS input files
Convert the training data to LAMMPS format:
conda activate /usr/local/usrapps/$GROUP/$USER/python_base
cd /share/$GROUP/$USER/lammps/nequip_source/configs
python -c "
from ase.io import read, write
atoms = read('fcu.xyz', index=0)
write('test.data', atoms, format='lammps-data', atom_style='atomic')
print('Species:', set(atoms.get_chemical_symbols()))
print('Num atoms:', len(atoms))
"
ASE maps the species alphabetically, so the atom types are 1=C, 2=Cu, 3=H, 4=O.Create the LAMMPS input file "in.test" in /share/$GROUP/$USER/lammps/nequip_source/configs
units metal atom_style atomic boundary p p p newton off read_data test.data mass 1 12.011 mass 2 63.546 mass 3 1.008 mass 4 15.999 pair_style nequip pair_coeff * * compiled_model.nequip.pth C Cu H O thermo 10 thermo_style custom step temp pe ke etotal press velocity all create 300.0 12345 fix 1 all nvt temp 300.0 300.0 0.1 timestep 0.001 dump 1 all xyz 100 trajectory.xyz run 1000
Step 6 — Run the MD simulation
Check the idle GPUs again and set --gres to match:si --gpus --qos short_gpu # Note, GPUs in other QOS (e.g., the "gpu" QOS) can be checked too, but for short jobs (less than 2 hours) it is generally advisable to use the high priority short_gpu QOS.Create "run_lammps.sh" in /share/$GROUP/$USER/lammps with the following lines:
#!/bin/bash #SBATCH --job-name=lammps_nequip #SBATCH --partition=gpu_partners #SBATCH --gres=gpu:h200:1 #SBATCH --qos=short_gpu #SBATCH --ntasks=1 #SBATCH --cpus-per-task=4 #SBATCH --time=00:10:00 source ~/.bashrc module load PrgEnv-intel/2025.3-slurm module load cuda/12.0 conda activate /usr/local/usrapps/$GROUP/$USER/python_base cd /share/$GROUP/$USER/lammps/nequip_source/configs srun /usr/local/usrapps/$GROUP/$USER/build_lammps/lammps/build/lmp -in in.testRun this batch job with:
sbatch run_lammps.shThe job produces:
- slurm-<jobid>.out — thermo output and performance statistics
- log.lammps — the same thermo output
- trajectory.xyz — atom positions every 100 steps, for visualization
Last modified: November 04 2026 12:29:47.
HPC Assistant