v0.6.0

Highlights

• New ORB-v3 backend (orb_v3_direct_omol, orb_v3_conservative_omol) alongside Fairchem UMA.
• DFT-D3(BJ) dispersion correction available for both backends.
• All four torch-sim optimizers selectable: FIRE, gradient_descent, LBFGS, BFGS.
• Upgraded to torch-sim 0.6.0 (InFlightAutoBatcher).

Configuration

• Restructured config (model, optimization, technical groups).
• New knobs: max_memory_padding, steps_between_swaps, memory_scaling_factor, max_atoms_to_try.
• Improved per-GPU device selection (cuda:N).
• Production defaults: memory_scaling_factor=1.75, max_memory_padding=0.95, steps_between_swaps=1.

Tests & docs

• Test suite expanded with per-backend (Fairchem/ORB) and optimization-focused suites.
• Documentation refreshed for ORB, D3, and the new configuration layout (online at https://niklashoelter.github.io/gpuma/).
• README badge bar (CI / PyPI / Python / docs / license).

Compatibility

• Python: 3.12 only (requires-python = "==3.12.*"). 3.11 and 3.13 are blocked by upstream pinned dependencies.