Backends¶
mliport runs exactly one backend per calculation. The backend comes from
--model-type, from a model alias or profile in your settings, or the run stops
with No MLIP backend was selected. Nothing falls back to something else.
| Backend | Page | --model-type |
Checkpoint used in the examples |
|---|---|---|---|
| MACE | MACE | mace |
mace-omat-0-medium.model |
| DPA (DeepMD-kit) | DPA | dpa |
DPA-3.1-3M.pt with --head Omat24 |
| GRACE | GRACE | grace |
GRACE-2L-OMAT-medium-base |
| UMA (FAIRChem) | UMA | uma (or fairchem) |
uma-s-1p2.pt with --task omat |
| NEP | NEP | nep (+ required --nep-runtime ase|gpumd) |
your own nep.txt |
Each backend lives in its own environment (.venv-mace, .venv-dpa,
.venv-grace, .venv, .venv-nep) because their frameworks and Python
requirements differ;
see Installation for how that works. To check that one of
them can really run, ask doctor:
DPA and GRACE need their GPU chosen before the framework loads. The command line
handles this by restarting itself in a fresh process with CUDA_VISIBLE_DEVICES
set; from the Python API you have to set it in the calling process yourself.
Every page below follows the same outline: how the backend is installed, which model profiles mliport has pinned, how to run it, and what to watch out for. The environment and profile tables are generated from the compatibility registry and the validation manifest, so they cannot drift from the code.
At a glance¶
| MACE | DPA | GRACE | UMA | |
|---|---|---|---|---|
| Architecture | equivariant message passing with many-body features | message passing on a line-graph series (pairs, angles, ...) | graph atomic cluster expansion (1L local, 2L semilocal) |
eSEN network with a mixture of linear experts |
| Example checkpoint trained on | OMat24 | OpenLAM-v1 (31 datasets) | OMat24 | OMat24, OMol25, OC20, ODAC25, OMC25 |
| Reference level of the example | PBE / PBE+U, bulk | per head; Omat24 = PBE / PBE+U |
PBE / PBE+U, bulk | per task; omat = PBE / PBE+U |
| You choose | precision (float64 / float32) |
--head |
the exported build | --task |
| Framework | PyTorch | PyTorch (TensorFlow for .pb) |
TensorFlow | PyTorch |
| License of the example | check the checkpoint | check the checkpoint | check the checkpoint | non-commercial research |
None of these checkpoints includes dispersion, and only UMA was trained on surfaces and molecules as well as bulk. A foundation model reproduces its reference functional, not experiment, and how well it does that on your system is something to check with a few DFT single points. The Models page explains how to compare backends honestly.