Changelog

Unreleased changes

0.3.1

  • get_upet now returns a non-scripted instance of AtomisticModel, allowing you to call AtomisticModel.save() on it.

0.3.0

  • Added the PET-MAD v1.6.0 models (sizes XS, S and M), trained on the MAD-1.6 dataset, which extends MAD-1.5 with catalytic surfaces. These are now the recommended PET-MAD models, and the ones resolved by version="latest".

  • Added pet-mad-m to the list of available models.

  • Added the UPETWrapper interface to the NVIDIA ALCHEMI Toolkit, for GPU-native batched inference, relaxations and MD. torch.compile support requires a 'solver' adaptive-cutoff checkpoint, i.e. PET-MAD v1.6.0 or newer.

  • Added force and stress uncertainties and ensembles to UPETCalculator.

  • Added an uncertainty_threshold option to UPETCalculator, which warns when the predicted atomic energy uncertainty exceeds the given value.

  • Model metadata (name, description, authors, references) is now read from the checkpoint when it provides it.

  • Updated to metatrain v2026.4 and fixed uncertainty estimation under rotational averaging.

  • Relaxed the nvalchemi-toolkit-ops and warp-lang version pins, and switched to deriving the package version from git tags with setuptools-scm.

  • Fixed the Gaussian filter used by the PET-MAD DOS calculator, which was applied incorrectly to batched inputs, and added a test covering it.

0.2.5

  • Updated the PET-MAD DOS calculator by adding a DOS denoising step, DOS alignment function, and other technical improvements.

  • Updated the docs for the new DOS calculator API.

0.2.4

  • Pin the version of warp-lang in dependencies as a hotfix around an internal change.

0.2.3

  • Added sphinx documentation for the package.

  • Switched to using metatomic_ase directly instead of metatomic.torch.ase_calculator.

  • Updated nvalchemi-toolkit-ops dependency to v0.3.0.

0.2.2

  • Added list_upet() function for listing available models.

  • Pinned nvalchemi-toolkit-ops dependency version.

  • Removed unused scipy dependency.

  • Enabled Windows CI tests.

0.2.1

  • Added pet-mad-xs to the available models list.

  • Added citation for PET-MAD-1.5.

0.2.0

  • Added PET-MAD v1.5.0 model.

  • Deprecated the PET-MAD and PET-OMAD models trained on the original MAD-1 dataset.

0.1.2

  • Added speed benchmarks for the UPET models

  • Added a support for loading the local checkpoints

  • Added variants selection for running different heads of the model

  • Modified the internal TorchScript exporting mechanism, so it saves a temporary model file on the disk.

0.1.1

  • Fixed a few bugs in the code, improved documentation and added a support for specifying the “latest” version of the model in the upet.get_upet and upet.save_upet functions.

0.1.0

  • Initial release of the UPET package, succeeding PET-MAD. Updated package structure, calculator names, and model naming conventions to reflect the new UPET branding and functionalities.

  • Added a support for new models trained on popular datasets for atomistic machine learning.