What’s Changed

New Features / Enhancements

• fix(compile): Remove graph break in ChgSpinEmbedding rand_emb path (#1907) @misko

Bug Fixes

• resolve small edge case bugs (#1920) @mshuaibii
• Fix Triton import failed torch-sim (TS) CI. Clean TS interface to avoid for loop and ase. (#1901) @CompRhys
• Fix permute_wigner_inv bwd_dw kernel add more than 128 channel support (#1902) @misko

What’s Changed

New Features / Enhancements

• md runner abstractions + improvements (#1835) @mshuaibii

What’s Changed

• Turbo mode by default now gets the UMA 1.2 speed up (previously non default behind execution backend uma_fast_gpu).
• Running turbo mode with MD is now alot smoother and less slowdown from torch recompiles

New Features / Enhancements

• fix nvalchemi version (#1905) @misko
• Consolidate turbo and turbo_umas (#1898) @rayg1234
• Optimize md torch compile (#1892) @misko
• Reduce at higher precision (#1889) @rayg1234
• Upstream TorchSim Interface to prevent CI issues and slippage with fa… (#1895) @lbluque
• Default untrained predictions (#1883) @lbluque
• checkpoint surgery script (#1858) @misko

Bug Fixes

• single atom bug fix and test (#1882) @wood-b

Dependencies

• numpy 2.4 (#1862) @lbluque

What’s Changed

New Features / Enhancements

• Enable Untrained Property Predictions Including Hessians (#1811) @lbluque

What’s Changed

This release adds ability to predict stress outputs for UMA 1.2 and reduces recompilation issues for the new UMASFastGPU backend due to bugs in torch compile.

• Refactor MLP Energy/EFS Heads with Shared Helper Functions (#1734) @lbluque

New Features / Enhancements

• Enable Untrained Property Predictions Including Hessians (#1811) @lbluque
• Add Hessian matrix calculation support (#1735) @lbluque
• multi-node profiling script (#1845) @rayg1234

What’s Changed

• UMA-S 1.2 Release!

New Features / Enhancements

• UMA-S 1.2 (#1861) @misko
• Refactor single atoms (#1732) @levineds
• add model id (#1859) @misko
• feat: add MOLE model support to DatasetSpecificMoEWrapper (#1848) @misko
• Add UnifiedRadialMLP for batched radial computation (#1831) @misko
• Umas fast gpu backend (#1826) @misko
• enbable GP for nvidia graphgen (#1791) @rayg1234
• [adsorbml] custom adsorbate support (#1822) @mshuaibii

Bug Fixes

• Disable chan balance compile (#1863) @rayg1234
• Revert 2 commits (#1860) @rayg1234
• Update inference settings (#1857) @wood-b
• Fix data loading imports in benchmarks (#1850) @lbluque
• Add workers arg to from_model_checkpoint (#1844) @rayg1234
• Quaternion-based Wigner D method for fixing y-aligned edges (#1771) @levineds
• Fix omc-is2re-polymorph reducer (#1837) @lbluque
• Revert graphgen v2 pre partitioning (#1838) @rayg1234
• add preemption state for runners that don't create it (#1834) @mshuaibii
• Fix per atom MAE bug (#1825) @EricZQu

Tests

• use uma-1p1 instead of uma-1 in tests (#1856) @lbluque

What’s Changed

New Features / Enhancements

• Generic MD Runner (#1782) @mshuaibii
• nvidia graph gen support (#1737) @mshuaibii
• add fp64 inference setting (#1789) @levineds
• Execution backend (#1802) @misko
• embeddings update (#1765) @wood-b
• composition dropout (#1767) @wood-b
• Fuse the edge envelope into wigner_mapping_inv instead of applying on each layer (#1792) @misko
• channel charge and spin (#1768) @wood-b

Bug Fixes

• Fix flaky unit tests (#1805) @levineds
• bug fix (#1812) @wood-b
• [ray] unique dashboard + worker ports for multiple local clusters (#1787) @mshuaibii

Dependencies

• numpy 1.26+ compatibility (#1738) @bkmi

What’s Changed

Bug Fixes

• Fix lammps cell (#1772) @rayg1234
• Fix lammps extract_atom inconsistency (#1703) @hsulab

Dependencies

• Add tests for python<=3.13 (#1600) @zulissimeta

What’s Changed

New Features / Enhancements

• local error handler (#1775) @mshuaibii
• Increase test pg timeout (#1788) @rayg1234
• Refactor FAIRChemCalculator to make it model-agnostic (#1731) @levineds
• Refactor MLIPPredictUnit be model-agnostic (#1727) @levineds
• load existing heads for finetuning (#1766) @wood-b
• Added ODAC25 models to pretrained.json and a package for odac data (#1719) @anuroopsriram
• Change ray launcher to run on a single slurm job (#1723) @rayg1234
• Use number of atoms for batch size (#1690) @lbluque

Bug Fixes

• move to file initialization for tests (#1785) @mshuaibii
• Fix NPT + GP + Mole bug (#1780) @rayg1234
• patch until torchtnt resolves install (#1763) @mshuaibii
• update sph_harm to sph_harm_y (#1736) @bkmi
• Fix port conflict in gp tests (#1745) @rayg1234

Documentation

• Jupyterbook>=2 docs refactor (#1627) @zulissimeta
• Fix typo in docstring for get_neighbors_pymatgen (#1715) @thomasloux

Dependencies

• bump ray[serve] (#1724) @lbluque

What’s Changed

Lots of Parallel and Batch inference upgrades:

• Added new highly efficient batch inference integration with Ray
• Graph building is now parallelized, UMA-s runs real MD with 1 ns/per day on 100k+ atoms with GP

New Features / Enhancements

• Early Partition and Parallel Graph Gen (#1630) @rayg1234
• Make a checkpointable Ray launcher (#1666) @rayg1234
• Batch server unit (#1622) @lbluque
• Formation energy calculator (#1538) @lbluque

Documentation

• UMA catalysis tutorial (#1667) @zulissimeta

Dependencies

• Add ray as dependency (#1675) @lbluque
• upgrade numpy/numba (#1537) @lbluque
• allow newer lmdb versions (#1655) @lbluque