Changelog#

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

Unreleased #

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

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

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.

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)

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.

TorchScript bindings to all metatensor-core class: Labels, LabelsEntry, TensorBlock, and TensorMap;
Implementation of mts_array_t/metatensor::DataArrayBase for torch::Tensor;

Expose TorchScript classes to Python;
Expose all functions from metatensor-operations as TorchScript compatible code;