metatomic-torchsim-v0.1.2

This release removes the upper-version pin on torch-sim-atomistic to make updating the code in there that re-exports this package easier.

metatomic-torchsim v0.1.1

This is a bugfix release, making sure the metatomic-torchsim wheel on PyPI properly declares its dependencies.

metatomic-torchsim v0.1.0

This is the first release of metatomic-torchsim, a standalone package containing the TorchSim integration for metatomic models.

This release adds the following features on top of the previously existing code from TorchSim:

• Support for output variants via the variants parameter, matching the ASE calculator's variant selection
• Non-conservative forces and stresses via non_conservative=True, reading model outputs directly instead of autograd
• Per-atom energy uncertainty warnings via uncertainty_threshold, triggered when the model provides energy_uncertainty with per_atom=True
• additional_outputs parameter for requesting arbitrary extra model outputs

metatomic-ase v0.1.0

This is the first release of metatomic-ase, a new package containing the ASE integration for metatomic models. This is the new home of the code that used to live in metatomic.torch.ase_calculator.

metatomic-torch v0.1.11

This is a patch release of metatomic-torch, adding the following feature:

The ASE calculator now supports nvalchemi-toolkit-ops as a backend for the calculation of full neighbor lists on CUDA devices. It will be used if available in the environment.

metatomic-torch v0.1.10

This release fix a bug introduced in v0.1.9 that made the ASE calculator unable to process mixed periodic and non-periodic boundary conditions.

metatomic-torch v0.1.9

This release introduces the following changes:

• The ASE calculator now always uses vesin to compute neighbor lists, including the ability to directly compute them on CUDA GPUs.
• pick_device properly handle explcit device index (i.e. selecting device="cuda:2")
• Removed support for torch 2.1 and 2.2

metatomic-torch v0.1.8

This is a new release of metatomic-torch, with the following changes:

Added

• Added the ability for models to request extra inputs from the simulation engine, on top of positions/cell/types. These extra inputs are requested by the model and stored on the system by the engine (like for neighbor lists). They are provided as TensorMap, stored in the system's data, and follow the same metadata structure as the standard outputs (anything that can be an output can also be an input).
• Added support for charges, masses, velocities as standard model inputs and outputs
• Added support for PyTorch v2.10

Changed

• Properties in standard output/inputs should now consistently use singular (i.e. position instead of positions). The plural names are deprecated and will be remove in the future.

metatomic-torch v0.1.7

This is a small patch release for metatomic-torch, with the following changes:

• Fixed metatomic.torch.ase_calculator.MetatomicCalculator to work even when a model does not have an energy output (#121)
• Added support for loading TorchScript extensions from deepmd-kit (#98)

metatomic-torch v0.1.6

Added

• Added support for torch v2.9
• ModelOutput now has a description field, to carry more information about a given output.
• Added the pick_output function that can be used by simulation engines to pick the correct output based on what's available inside a model and which variant (if any) the user requested.
• Added metatomic.torch.ase_calculator.SymmetrizedCalculator, an ASE calculator that can average energy calculations over the O(3), SO(3) or space group of a crystal, to make unconstrained, non-equivariant models rotationally equivariant up to the integration order.

Changed

• The pick_device function now returns a torch::DeviceType instead of a string in C++. This does not affect the Python API.
• It is now possible to construct an AtomisticModel with a torch module already compiled with TorchScript.

Fixed

• Variants can now be used with non-standard outputs as well (#105)

Removed

• We dropped support for Python 3.9, and now requires at least Python 3.10