xnn.common.models.hub.loader.from_pretrained#

xnn.common.models.hub.loader.from_pretrained(source, *, wrap=True, **kwargs)[source]#

Load a pre-trained model in one call.

MACE foundation models, xnn-trained models, Zenodo uploads and local checkpoints all load the same way:

from xnn.common.models import from_pretrained

model = from_pretrained("mace-off23-small")             # registry name
model = from_pretrained("mace-mh-0", head="omat_pbe")   # one head
model = from_pretrained("doi:10.5281/zenodo.18957344",
                        filename="mace_csfapbbri_al_5_1_stagetwo.model")
model = from_pretrained("runs/exp/best.pt")             # trainer checkpoint
model = from_pretrained("mace-off23-small", cache_dir="/scratch/models")

The first call downloads (and verifies) the file into cache_dir and, for a foreign format such as mace-torch, converts it once; later calls load the cached portable directory without network access and without the upstream package.

Parameters:

source (str or pathlib.Path) – Registry name, local directory or checkpoint, URL, or Zenodo DOI; see fetch_model().
wrap (bool, optional) – Return the ForceStressOutput wrapper (forces, and stress with compute_stress=True). False returns the bare potential inside it. Defaults to True.
**kwargs – cache_dir, filename, head, format, force_download, local_files_only, quiet, device, dtype, dispersion, compute_forces, compute_stress, eeq_reuse, use_fast, model_options; see load_pretrained().

Returns:

The model in eval mode.

Return type:

torch.nn.Module