Changelog

All notable changes to metatensor-torch are documented here, following the keep a changelog format. This project follows Semantic Versioning.

Unreleased

Version 0.8.1 - 2025-09-24

Fixed

• The custom metatensor_torch::Module now properly handles nested containers (list, dict, tuples) with empty containers inside (#978)

Version 0.8.0 - 2025-08-22

Removed

• The metatensor.torch.atomistic python module now lives in the metatomic-torch package (https://github.com/metatensor/metatomic); as well as the corresponding C++ code (#include <metatensor/torch/atomistic.hpp>).

Added

• Added support for PyTorch v2.8

• A custom C++ class metatensor_torch::Module that should be used instead of torch::jit::Module when the modules contains metatensor data (Labels, TensorBlock, TensorMap) as attributes. This class will properly handle moving this data to the correct dtype and device when calling module.to(). The corresponding Python class is defined in metatensor-learn.

• There is now an assume_unique flag in the Labels constructor. This allows bypassing a check when the user can ensure beforehand that all entries will be unique.

Changed

• TensorMap.to, TensorBlock.to, and Labels.to now accept a non_blocking argument, with the same meaning as in torch.Tensor.to.

• metatensor.torch.save, metatensor.torch.load, metatensor.torch.load_block and metatensor.torch.load_labels can now take pathlib.Path objects as inputs, as well as pre-opened files. When using these functions from TorchScript, only str arguments are supported.

Version 0.7.6 - 2025-04-25

Added

• Added support for PyTorch v2.7

Changed

• The lowest supported version of PyTorch was increased from v1.12 to v2.1

Version 0.7.5 - 2025-04-10

Added

• Labels.difference and Labels.difference_and_mapping for finding the set difference between two Labels

• energy_uncertainty, non_conservative_forces and non_conservative_stress are now possible outputs for atomistic models

• The ASE calculator can now use non_conservative_forces and non_conservative_stress outputs, which are selected with the non_conservative flag

Version 0.7.4 - 2025-03-28

Added

• Added MetatensorCalculator.compute_energy() to provide batched calculation of energies, forces and stresses (#880).

Changed

• MetatensorCalculator.run_model() now supports batching across multiple ase.Atoms objects (#879).

Version 0.7.3 - 2025-02-19

Changed

• The MetatensorCalculator for ASE now only computes the stress if the system is periodic in all directions

Version 0.7.2 - 2025-02-18

Fixed

• Fixed a bug with the custom TorchScript extensions when using featomic-torch

Version 0.7.1 - 2025-02-17

Changed

• The default extension for file serialization is now .mts instead of .npz

Added

• Support for Python 3.13 and PyTorch 2.6

Version 0.7.0 - 2025-01-22

Changed

• Custom data stored in System is now stored as TensorMap instead of TensorBlock. This allow for more kinds of data to be stored.

• metatensor_torch::TorchLabels is now metatensor_torch::Labels; and similarly for metatensor_torch::TensorBlock and similarly for metatensor_torch::TensorMap. We temporarily provide type aliases for backward compatibilty.

Added

• System can now be serialized and saved to a file using metatensor.torch.atomistic.save and metatensor.torch.atomistic.load_system. For the initial implementation, these functions are only accessible from Python.

Version 0.6.3 - 2024-12-11

Fixed

• Fixed an issue with Python wheels building, where the underlying C++ library would get a different version from the Python library and fail to load.

Version 0.6.2 - 2024-12-11

Fixed

• Selecting from empty Labels no longer fails (#792)

• Models are exported with requires_grad=False for all parameters, which should make them faster when using autograd to compute forces and stress (#802)

Version 0.6.1 - 2024-11-7

Fixed

• Added missing data to NeighborListOptions serialization (#784)

Changed

• load_atomistic_model now returns a MetatensorAtomisticMode instead of a raw TorchScript model. This allows to reload, modify and re-export a model (#783).

Version 0.6.0 - 2024-10-29

Added

• a is_atomistic_model (Python only) function to check if an loaded module is an metatensor atomistic model (#697, #766)

Changed

• the System class now supports boundary conditions along some axes but not others. This is implemented via a new pbc attribute. Any non-periodic dimension in a System must have the corresponding cell vector set to zero.

• NeighborListOptions can now request strict or non-strict neighbor lists (#770)

Version 0.5.5 - 2024-09-03

Added

• a "features" standard output for atomistic models (#718)

Fixed

• the Python wheels request the right versions of torch in their metadata (#724)

Version 0.5.4 - 2024-08-28

Added

• read_model_metadata to load only the ModelMetadata from an exported atomistic model without having to load the whole model.

• Users can now store arbitrary additional metadata in ModelMetadata.extra

• Added Labels.select function to sub-select entries in labels

• Added support for serialization of TensorBlock with TensorBlock::load, TensorBlock::load_buffer, TensorBlock::save, TensorBlock::save_buffer and the corresponding functions in metatensor.torch.

Version 0.5.3 - 2024-07-15

Changed

• MetatensorAtomisticModel.save() always saves models on the CPU.

• We now require Python >= 3.9

Fixed

• Fixed a memory leak in register_autograd_neighbors (#684)

Version 0.5.2 - 2024-06-21

Added

• MetatensorAtomisticModel.save() to save a wrapped model to a file.

• TensorBlock.__len__ and TensorBlock.shape, which return the length and shape of the values in the block respectively (#640)

• metatensor.torch.atomistic.ase_calculator.MetatensorCalculator can now use vesin for faster neighbor list calculations (#659)

• When running atomistic models in the PyTorch profiler, different sections of the code now have meaningful names

Deprecated

• MetatensorAtomisticModel.export() is deprecated in favor of MetatensorAtomisticModel.save()

Fixed

• metatensor.torch.atomistic.ase_calculator.MetatensorCalculator uses the right device when computing stress/virial (#660)

Version 0.5.1 - 2024-05-14

Added

• preprocessor macros containing the version number of metatensor-torch: METATENSOR_TORCH_VERSION, METATENSOR_TORCH_VERSION_MAJOR, METATENSOR_TORCH_VERSION_MINOR, and METATENSOR_TORCH_VERSION_PATCH.

Version 0.5.0 - 2024-05-02

Changed

• We renamed neighbors_list to neighbor_list in all functions (#587)

metatensor-torch Python

Changed

• The neighbor lists calculation in MetatensorCalculator (ASE calculator based on metatensor models) is now a lot faster (#586)

• Multiple small improvements related to custom TorchScript extensions (#584)

• There are reference output for neighbor list calculations, which should help writing interfaces to metatensor models in new simulation engines (#588)

• The wheels for metatensor-torch on PyPI now declare which versions of torch they are compatible with (#592)

Version 0.4.0 - 2024-04-11

metatensor-torch C++

Added

• ModelCapabilities::dtype, used by the model to communicate the dtype it wants to use for inputs and outputs.

• The load_model_extensions() function to facilitate loading a model using TorchScript extensions.

Changed

• System::add_data now has an override parameter to replace custom data with a new value.

metatensor-torch Python

Changed

• We now release wheels compatible with multiple torch versions on PyPI, removing the need to compile C++ code when installing this package.

Added

• ModelCapabilities.dtype, used by the model to communicate the dtype it wants to use for inputs and outputs.

• The device that should be used by a model inside the ASE’s MetatensorCalculator can now be specified by the user.

• The load_model_extensions() and load_atomistic_model functions to facilitate loading a model using TorchScript extensions

Version 0.3.0 - 2024-03-01

metatensor-torch C++

Added

• ModelMetadata to record metadata about a specific model such as it’s name, authors, etc.

• Added interaction_range and supported_devices to ModelCapabilities

Changed

• System::species has been renamed to System::types.

metatensor-torch Python

Added

• ModelMetadata to record metadata about a specific model such as it’s name, authors, etc.

• Added interaction_range and supported_devices to ModelCapabilities

Changed

• System.species has been renamed to System.types.

Version 0.2.1 - 2024-01-26

metatensor-torch C++

Added

• Offer serialization functionality as member functions (i.e. TensorMap::load) in addition to the existing free standing functions (i.e. metatensor_torch::load) (#453)

• In-memory serialization with TensorMap::save_buffer, TensorMap::load_buffer, and the respective free standing functions (#455)

• Serialization of Labels, with the same API as TensorMap (#455)

metatensor-torch Python

Added

• Offer serialization functionality as member functions (i.e. TensorMap.load) in addition to the existing free standing functions (i.e. metatensor.torch.load) (#453)

• In-memory serialization with TensorMap.save_buffer, TensorMap.load_buffer, and the respective free standing functions (#455)

• Serialization of Labels, with the same API as TensorMap (#455)

Version 0.2.0 - 2024-01-08

metatensor-torch C++

Added

• New classes specifically tailored for atomistic models (#405):

  • System defines the input of a model;

  • NeighborsListOptions allow a model to request a specific neighbors list;

  • ModelRunOptions, ModelOutput and ModelCapabilities allow to statically describe capabilities of a model, and request specific outputs from it.

• TensorBlock::to, TensorMap::to, and System::to to change the device or dtype of torch Tensor stored by metatensor

• Labels::device, TensorBlock::device and TensorMap::device; as well as TensorMap::scalar_type, and TensorBlock::scalar_type to query the current device and scalar type/dtype used by the data.

• metatensor_torch::version function, returning the version of the code as a string.

Fixed

• We now check that all tensors in a TensorBlock/TensorMap have the same dtype and device (#414)

• keys_to_properties, keys_to_samples and components_to_properties now keep the different Labels on the same device (#411)

metatensor-torch Python

Added

• New classes specifically tailored for atomistic models (#405):

  • same classes as the C++ interfaces, in metatensor.torch.atomistic

  • MetatensorAtomisticModel as a way to wrap user-defined torch.nn.Module and export them in a unified way, handling unit conversions and metadata checks.

• ASE calculator based on MetatensorAtomisticModel in metatensor.torch.atomistic.ase_calculator. This allow using arbitrary user-defined models to run simulations with ASE.

• TensorBlock.to, TensorMap.to and System.to to change the device or dtype of torch Tensor stored by metatensor

• Labels.device, TensorBlock.device and TensorMap.device; as well as TensorMap.dtype, and TensorBlock.dtype to query the current device and dtype used by the data.

Version 0.1.0 - 2023-10-11

metatensor-torch C++

Added

• TorchScript bindings to all metatensor-core class: Labels, LabelsEntry, TensorBlock, and TensorMap;

• Implementation of mts_array_t/metatensor::DataArrayBase for torch::Tensor;

metatensor-torch Python

Added

• Expose TorchScript classes to Python;

• Expose all functions from metatensor-operations as TorchScript compatible code;