TMol 0.3.0 adds Apple Silicon GPU acceleration, improves CPU/CUDA packing and relaxation, and simplifies installation and CI.
Changes since 0.2.0:
- Apple Metal offload accelerates scoring and packing while poses remain on the CPU. Set
TMOL_MPS_OFFLOAD=0for CPU-only execution. - Updated public APIs for structure I/O, scoring, packing, and relaxation, with revised tutorials. FastRelax selects Cartesian or kinematic minimization through
CartesianRelaxMinimizerandKinematicRelaxMinimizer; replace the oldmin_fn=argument withminimizer=. - Faster CPU/CUDA FastRelax, configurable annealing trajectory counts, CUDA graph support for constrained relaxation, and reduced CUDA cache growth.
- Pose selection, repetition, and chain-ID helpers; kinematics and line-search improvements; corrected backbone restraints and bond-range handling.
- Slimmer direct dependencies. Install
tmol[ligand]for OpenBabel-dependent ligand preparation and noncanonical chemistry, or combine it with CUDA, for exampletmol[cu13,ligand]. Development and documentation extras include ligand tools. - Routine CPU CI runs on GitHub-hosted Linux x86-64, Linux ARM64, and macOS ARM64. Maintainers can dispatch CUDA CI; release publication continues to require GPU wheel scoring.
- BSD-3-Clause licensing, vendored license notices, and repository/documentation links under RosettaCommons.
Installation:
python -m pip install tmol==0.3.0
# With an existing matching CUDA-enabled PyTorch:
python -m pip install 'tmol[cu13,ligand]==0.3.0'Python 3.11–3.14 and PyTorch 2.11+ are supported. Native wheels cover Linux x86-64/ARM64 and Apple Silicon. CUDA 12 and CUDA 13 operators are separate optional packages at the same TMol version. See the installation guide for platform and driver requirements: https://rosettacommons.github.io/tmol/latest/installation.html
Validated with AtomWorks 3.0.4, the latest public PyPI release at publication time. The Biotite 1.6.0 pin matches its structure I/O requirements; AtomWorks has no upper version bound in TMol.