QCMLForge pretrained models

Pretrained checkpoints for QCMLForge (apnet_pt). Files are resolved on demand by apnet_pt.hf_pretrained; set QCMLFORGE_AUTO_DOWNLOAD_PRETRAINED=1 to allow non-interactive downloads.

APNet2 weight sets

weights=	files	description
qcmlforge (default)	am_ensemble/am_{0..4}.pt, ap2_ensemble/ap2_{0..4}.pt	APNet2 ensemble trained by QCMLForge.
ap2_tf_paper	ap2_tf_paper/atom_models/atom{0..4}.pt, ap2_tf_paper/pair_models/pair{0..4}.pt	The AP-Net2 ensemble published with the paper (zachglick/apnet), converted from TensorFlow.

from apnet_pt.pretrained_models import apnet2_model_predict

pred = apnet2_model_predict(dimers, weights="ap2_tf_paper")

About ap2_tf_paper

These are the original TensorFlow SavedModel weights converted to PyTorch checkpoints, not a retrained approximation: every atom-model tensor transfers bit-exactly, and predictions match recorded TensorFlow output to ~1e-6 on multipoles and ~1e-4 kcal/mol on interaction energies (float32 reduction order).

Against the paper's reported MAEs, the five-member average matches exchange, induction and dispersion to 2-6e-4 kcal/mol and sits +0.134 kcal/mol high on electrostatics, which is still open; see the parity spec.

Two things they require, both handled by weights="ap2_tf_paper":

• each pair{i} must be loaded with its own atom{i} — the checkpoints carry no embedded atom submodel;
• quadrupole_scale must be 1.5 (it is stored in the checkpoint config, not the state dict; dropping it shifts electrostatics by ~0.5 kcal/mol with no error).

They also cannot be loaded by the fused APNet2_AM_MPNN model.

Details, the parity table, and the conversion procedure: docs/apnet2-tensorflow-weights.md.