Use a model

Every framework lives in its own env, so a model always runs under that env's interpreter.

Ask your LLM

Relax POSCAR with MACE-MPA-0 and give me the relaxed structure.

Write an MD script for this slab that uses SevenNet.

The agent takes the calculator lines from the registry and runs your script with the right interpreter. To do it yourself, there are two ways.

In your own script

Use this for MD, relaxations and anything long. Copy the lines once; your script does not need oh-my-mlip at run time.

import sys; sys.path.insert(0, "<repo>")       # or $OH_MY_MLIP_HOME
from oh_my_mlip import resolve

spec = resolve("MACE")                         # a family, or a version such as "MACE-MH-1-OMAT"
print(spec["python"])      # the interpreter to run your script with
print(spec["imports"])     # import lines
print(spec["inference"])   # the line that creates `calc`
print(spec["env_run"])     # environment variables to export first (often empty)

Paste imports and inference verbatim:

# my_run.py  ->  run with spec["python"]
from ase.io import read
from mace.calculators import mace_mp
atoms = read("POSCAR")
calc = mace_mp(model='medium-mpa-0', dispersion=False, default_dtype='float64', device='cuda')
atoms.calc = calc
print(atoms.get_potential_energy())

Inside that env you can also build the calculator in one call: oh_my_mlip.get_calculator("MACE").

From one script, across models

Use this for quick comparisons. Each model runs in its own env process, so the calling Python needs no model installed.

import oh_my_mlip

out = oh_my_mlip.run("MACE", atoms)            # {"energy": ..., "forces": ...}

with oh_my_mlip.WorkerPool() as pool:          # workers stay alive between calls
    for name in ("MACE", "SevenNet"):
        resp = pool.request(name, atoms)
        print(name, resp["results"]["energy"])

Every call crosses a process boundary, so use the first way for MD loops.

Ready-made scripts

From the repository root:

python run_examples/single_point.py MACE --structure POSCAR
PY=$(python -c 'from oh_my_mlip import resolve; print(resolve("MACE")["python"])')
"$PY" run_examples/relax.py MACE --structure POSCAR    # writes relaxed.extxyz

D3 dispersion

oh_my_mlip.run(..., apply_d3=True), or inside the model's env:

from catbench.dispersion import DispersionCorrection
calc = DispersionCorrection().apply(calc)

NequIP and Allegro

Their calculators load a model compiled for your GPU architecture during install. To prepare jobs for a different GPU, pass its architecture: resolve("NequIP", arch="sm86"). See GPU-architecture compilation.

Full procedure: recipes/run.md · per-framework lines and weights: Environment recipes.