This repository contains scripts and examples for building LAMMPS with the ML-IAP and NequIP pair styles, preparing NequIP foundation models for LAMMPS, and running MD benchmarks.
Setup and build scripts for a micromamba environment with NequIP dependencies and for compiling LAMMPS with the ML-IAP pair style and Kokkos (CUDA, Ampere80).
| File | Description |
|---|---|
| env_setup.sh | Instructions (meant to be run interactively on CPU) for creating a micromamba env, installing Python 3.11, PyTorch (CUDA 12.9), NequIP, and LAMMPS, and applying the pair_nequip_allegro patch. |
| build.sh | SLURM build script that loads modules, activates the micromamba env, and runs CMake. Builds the lmp executable in build/ |
Scripts to download the NequIP-L foundation model and compile it for use with LAMMPS. It needs to be run on the same hardware + environment as the final runs.
| Script | Purpose |
|---|---|
| compile_for_baseline.sh | Downloads the model and produces two variants under compiled_models/: (1) nequip-prepare-lmp-mliap --no-compile for the ML-IAP integration, and (2) nequip-compile --mode torchscript for the pair_style nequip TorchScript backend. |
| compile_for_test.sh | Same but for test purposes (without requesting the A100 80GB). |
Ensure the micromamba env path (e.g. ../build_lammps/env) and SLURM account/partition match your system before submitting.
Example LAMMPS input files and SLURM submission scripts for running NequIP models with Kokkos GPU.
| Path | Description |
|---|---|
| baseline_measurement/ | Baseline performance runs for the pair_style nequip and the TorchScript .pth from compile_for_baseline.sh; mliap-nocompile/ uses pair_style mliap unified with the .nequip.lmp.pt potential. |
| test/ | Example 1-GPU run for Si |
Update paths in the run scripts (lammps executable, micromamba env, and model paths in each lammps.in) to match your build_lammps and models locations.