88 lines (54 loc) · 2.66 KB

Changelog

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

Unreleased

Version 0.5.3 - 2026-03-10

Fixed

We now use the same cudart library from Python and C++ when using torch tensors with vesin.NeighborList.

Version 0.5.2 - 2026-02-26

Fixed

The results are now returned with the same dtype as the input even when we return an empty neighbor list

Version 0.5.1 - 2026-02-24

Added

Added vesin.metatomic.compute_requested_neighbors_from_options to allow computing all neighbor lists requested by a metatomic model in a TorchScript-compatible way.

Changed

Users can control how much memory is allocated for the CUDA neighbor list with the VESIN_CUDA_MAX_PAIRS_PER_POINT environment variable.
We now also try to load libcudart.so.{11,12,13} if the code can not find libcudart.so on Linux.

Version 0.5.0 - 2026-02-19

Added

vesin can now compute neighbor lists on GPUs using CUDA. This is automatically integrated in vesin-torch, and can be used with CuPy in the vesin python package.
vesin.NeighborList now accepts either numpy arrays, cupy arrays or torch tensor for the points, box, and periodic parameters.

Version 0.4.2 - 2025-11-06

Changed

arbitrary box orientations are now supported with mixed PBC (#88)

Version 0.4.1 - 2025-11-03

Added

vesin-torch wheels on PyPI now support PyTorch v2.9

Version 0.4.0 - 2025-10-27

Added

vesin now offers a Fortran API, you can enable it by giving the -DVESIN_FORTRAN=ON option to cmake (#50)
the single file build now contains a comment mentionning which version of vesin the file corresponds to.

Changed

The periodic argument to vesin_neighbors() in C and C++, NeighborList.compute() in all other languages can now be set separatly for each box dimension, allowing mixed periodic and non-periodic boundary conditions.
VesinDevice in C and C++ is now a struct containing both a device kind (i.e. CPU, CUDA, etc.) and a device index.