xnn.common.deploy.torchscript.export_torchscript_potential#

xnn.common.deploy.torchscript.export_torchscript_potential(model, cutoff, path, metadata=None, total_charge=0.0, spin_multiplicity=1.0)[source]#

Script a trained model to a standalone .pt and save it.

Wraps model in TorchScriptPotential, compiles it with torch.jit.script and writes the archive. The cutoff and the long_range / dispersion / charges flags are embedded as extra files, alongside any caller metadata, so a consumer can introspect the artifact without xnn.

Parameters:

model (torch.nn.Module) – The trained model (optionally LES-wrapped).
cutoff (float) – Neighbor-list cutoff radius to bake in.
path (str) – Destination .pt path.
metadata (dict or None, optional) – Extra key/value metadata; values are stringified.
total_charge (float, optional) – Net charge of the deployed system (D4 EEQ charges, charge-predicting cores), by default 0.
spin_multiplicity (float, optional) – Its spin multiplicity (two-channel AIMNet2 cores), by default 1.

Returns:

The path written.

Return type:

str