Release v0.4.0

0.4.0 - 2026-02-23

Summary

nvMolKit 0.4.0 adds GPU-accelerated substructure searching, optional stream control across Python APIs, and enhancements to Butina clustering.

Contributors

• Kevin Boyd (@scal444)
• Eva Xue (@evasnow1992)

Features

• GPU-accelerated substructure search with hasSubstructMatch, countSubstructMatches, and getSubstructMatches. Supports batch queries against batch targets with SMARTS-based query molecules.
• Optional stream parameter added to fingerprint generation, similarity, and Butina clustering APIs, enabling explicit CUDA stream control
• Butina clustering now supports optional centroid reporting via the return_centroids parameter (#82)
• Butina clustering performance improved by replacing CPU loops with CUDA Graph conditional nodes (#72)

Bug Fixes

• Fix data races when torch operations immediately followed nvMolKit calls on the default stream (Issue #84). Operations now correctly use the current stream or an explicit stream parameter (#36).
• Fix setup.py compatibility on some Python versions and rework CUDA target detection (#68)

Release v0.3.0

0.3.0 - 2025-12-12

Summary

nvMolKit 0.3.0 adds Butina clustering support, improved performance to MMFF relaxation and conformer generation, and increased compatibility with libraries and compilers.

Contributors

• Kevin Boyd (@scal444)
• Eva Xue (@evasnow1992)
• Xuangui Huang (@stslxg-nv)

Features

• Butina clustering API enabled, using distance matrix input. On an H200 GPU, speedups of 400-1000x can be achieved on datasets up to 60k molecules
• Improvements to BFGS minimizer. Up to 5x speedup compared to nvMolKit v0.2 on batches of small molecules (<20 atoms), with ~10-20% speedup in the general case. Applies to both MMFF relaxation and conformer generation.
• Conda-forge releases now support RDKit versions 2024.9.6 to 2025.9.3

Bug Fixes

• Fixed a bug where synchronizations on the wrong stream could lead to data races in tests (Issue #28)
• Fixed several areas where a memcpy could go out of scope before completing (Issue #28, Issue #29)
• Fixed a bug where ETKDG would exit early with small CPU counts due to an incorrect identification of resource mis-configuration (Issue #31)

Miscellaneous

• (C++) Added support for CUB/CCCL > v2.8
• (C++) Added support for externally specified CCCL
• (C++) Added support for CUDA 13.0

nvMolKit v0.2.0

0.2.0 - 2025-10-24

Summary

nvMolKit 0.2.0 comes with significant usability and feature-completeness improvements to existing functionality. It is also
the first release to have a conda-forge release.

Contributors

• Kevin Boyd (@scal444)
• Eva Xue (@evasnow1992)
• Ignacio Pickering (@IgnacioJPickering)

Features

• Add memory-segmented cross-similarity code, enabling larger datasets on systems with limited GPU memory (#13)
• Support conformer deduplication in ETKDG conformer generation (#14)
• Allow molecules > 256 atoms in conformer generation and MMFF optimization (#16)
• Enable all combinations of (ET)(K)(DG) in conformer generator (#17)

Bug Fixes

• Fix compilation error on C++ build with target=native on Hopper architecture GPUs. (#6)
• Fix lack of device-set cleanup in multi-GPU code (#8)
• Fix bug in fingerprint bool->bitfield packing/unpacking code (#11)
• Fix integer overflow leading to incorrect allocations in similarity calculation code. (#20)
• Fix crash in most multithreaded APIs whenever exceptions are thrown inside of OpenMP loop. Exceptions now properly propagated to python (#18)

Miscellaneous

• Removed unsupported Bulk Similarity APIs (#12)

v0.1.0

Initial release of nvMolKit