v0.5.1

Summary

nvMolKit 0.5.1 is a bug fix and quality-of-life release.

Contributors

• Kevin Boyd (@scal444)
• Eva Xue (@evasnow1992)
• Clay Moore (@mooreneural)

Bug Fixes

• Fix pip wheels failing to import in minimal manylinux/CUDA containers (#174).
• Fix incorrect TFD values for molecules with very large conformer sets and avoid invalid values for empty ring-torsion cases (d265234).
• Fix MMFF/UFF convergence parity with RDKit 2026.03+ for force fields with negative intermediate energies (rdkit/rdkit#9298).
• Fix empty Morgan fingerprints when every molecule in a batch exceeded the GPU size buckets (#195).
• Fix rare one-bit Morgan fingerprint mismatches in multi-round GPU batches (#197).
• Fix fused Butina clustering failures on very large inputs (#194).
• Fix ETKDG cases that incorrectly produced zero conformers for affected molecules (#202).
• Fix disconnected SMARTS queries raising the expected Python error instead of terminating the process with multithreaded preprocessing (#203).
• Fix recursive SMARTS queries with pip-installed RDKit builds (#208).
• Fix brittle Python import orders for MMFF and UFF optimization modules (#212).

Miscellaneous

• Clustering and similarity APIs now accept AsyncGpuResult, CUDA tensors, CPU tensors, and NumPy arrays consistently (#207).
• Autotune defaults now search fewer redundant batch sizes and avoid oversubscribing CPU threads (#179).