Python API

oh_my_mlip is imported from the repository path; it is not a pip package.

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

Importing it never loads a framework, torch or a GPU.

Registry

list_models() -> list[str] : Registered frameworks.

list_versions(model) -> list[str] : Versions of one framework.

resolve(model, version=None, *, arch=None) -> dict : The lines that use a model, without loading it. model is a framework ("MACE", resolving its default version) or a version ("MACE-MH-1-OMAT"). Main keys:

Key	Meaning
python	interpreter of the model's env
imports	import lines
inference	line(s) that create calc
env_run	environment variables to set when running
version	the resolved version
gated, license_url	whether the weights need a license and token
arch_pinned, arch	NequIP/Allegro: compiled model for a GPU architecture

Inside a model's env

get_calculator(model, version=None, device="cuda", apply_d3=False, *, arch=None) : Build the ASE calculator. Must run under that model's interpreter.

Across envs

run(model, atoms, properties=("energy", "forces"), device="cuda", apply_d3=False, *, version=None, arch=None) -> dict : Start the model's env process, compute once, stop it. Returns the results, e.g. {"energy": ..., "forces": ...}.

Worker(model, version=None, device="cuda", apply_d3=False, arch=None) : One persistent env process. start(), request(atoms, properties), shutdown(); usable as a context manager. request returns {"id": ..., "ok": True, "results": {...}} or {"ok": False, "error": ...}.

WorkerPool(device="cuda", apply_d3=False) : One worker per model, started on first use and restarted once if it crashes. request(model, atoms, properties, *, version=None), shutdown(); usable as a context manager.

Atoms are passed to worker processes as JSON; no pickle is used.