Changelog

Unreleased

  • Use local sorting and tree construction for one-device shard maps, avoiding unnecessary redistribution, domain adjustment and particle padding requirements.

Version 1.1.0

  • Added TreeHierarchy.poslvl() for accessing node positions and levels.

  • Added separate query/source indexing for interaction lists, including updated CUDA bindings. The Python interface and CUDA backend must be upgraded together.

  • Improved shard-map context detection, including single-device mapped contexts, and fixed multi-GPU edge cases.

  • Fixed integer dtype consistency with JAX x64 enabled and added validation and tests.

  • Explicitly blocked gradients through tree construction.

  • Fixed mass-centering for dimensions other than three.

  • Extended initial-condition helpers, statistics, and array utilities.

  • Configurations now use slots: assigning an unknown parameter raises an error instead of silently adding an unused attribute. Existing parameters remain mutable.

  • CUDA 12 release wheels now target manylinux_2_28 (Linux x86-64, glibc >=2.28), matching CUDA 13.

Version 1.0.3

  • Updated .vma access to support jax 0.10

  • Fixed issue with CUDA_MAJOR version detection for source builds

  • Added source links to documentation

Version 1.0.2

Initial stable release