Changelog#

0.6.0 - 2026-08-13#

Summary#

nvMolKit 0.6.0 adds two new GPU-accelerated capabilities: Maximum Common Substructure (MCS) search and a FIRE minimizer option for MMFF and UFF force field optimization. Fused Butina clustering has been rewritten on a pure CUDA backend, making it over 10x faster and removing the Triton dependency, and Butina clustering now supports reordering=False and matches RDKit output exactly. Substructure search is up to 2x faster end to end with a new DFS backend. This release supports RDKit 2025.09.1 through 2026.03.5 and includes all fixes from the v0.5.1 patch release.

Contributors#

  • Clay Moore (@mooreneural)

  • Eva Xue (@evasnow1992)

  • Kevin Boyd (@scal444)

  • Matthew Neba (@Matthew-Neba)

  • Timur Rvachov (@trvachov)

Features#

  • GPU-accelerated Maximum Common Substructure (MCS) search via the new nvmolkit.mcs.findMCS API. Searches many molecule pairs in one batched call — all pairs of a molecule set, an explicit pair list, or two zipped lists — with configurable atom/bond matching, per-pair timeouts, and a CPU RDKit fallback for oversized inputs. RDKit fMCS options that are not yet supported raise informative errors (#221)

  • FIRE minimizer as an alternative to BFGS for force field optimization, roughly 2x faster than BFGS for comparable accuracy. Available via minimizerKind="FIRE" in the MMFF and UFF optimization APIs and the BatchedForcefield API (#216)

  • Fused Butina clustering rewritten from Python/Triton onto a CUDA backend, keeping the full clustering loop on the GPU. Up to 10x faster and removes the Triton dependency, by @Matthew-Neba (#226)

  • Support for reordering=False in Butina clustering, matching the RDKit option, by @Matthew-Neba (#192)

  • Butina clustering results now match RDKit exactly, with deterministic tie-breaking and cluster ordering, by @Matthew-Neba (#241)

  • New DFS-based substructure search backend, up to 2x faster end to end, with reduced host-device transfer overhead (#244)

  • Conformer RMSD can now return a square distance matrix (output_format="square") for direct use with downstream clustering APIs (a26ae9c)

  • Performance improvements to conformer RMSD kernels by @mooreneural (ea09e2d)

  • Reduced per-call launch overhead in similarity computations by @mooreneural (ea09e2d)

Bug Fixes#

  • All bug fixes from the v0.5.1 patch release, covering ETKDG conformer generation, Morgan fingerprints, fused Butina clustering, TFD, SMARTS handling, MMFF/UFF convergence, and pip packaging

  • Fix a data race in batched MMFF minimization when a molecule’s conformers spanned multiple batches, which could produce incorrect atom typing or crashes at high thread counts (#236)

  • Fix MMFF failing to parametrize cyclophosphazine-type molecules (#258)

  • Fix an incorrect sixth-order torsion gradient term in distance geometry minimization (#217)

  • CUDA streams belonging to a different GPU are now rejected with a clear error instead of being silently accepted (#235)

Miscellaneous#

  • Document fixes for common torch CUDA-backend installation issues with pip, uv, and conda (#209)

  • New downloadable agent skill teaching LLM coding agents to use nvMolKit, with evals and an NVIDIA skills catalog card (434440f)

  • (C++) Major CMake / include-path refactor; all includes are now relative to the project root (#191)

0.5.1 - 2026-06-24#

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, by @mooreneural (d265234).

  • Fix MMFF/UFF convergence parity with RDKit 2026.03+ for force fields with negative intermediate energies, by @mooreneural (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).

0.5.0 - 2026-05-13#

Summary#

nvMolKit 0.5.0 adds three new GPU-accelerated APIs: Torsion Fingerprint Deviation (TFD), pairwise conformer RMSD, and UFF force field optimization. It also introduces a BatchedForcefield Python API for MMFF and UFF with constraints, custom options, and multi-conformer minimization; a low-memory fused Butina clustering path that avoids the O(N²) distance matrix; a Python autotuning framework for the main APIs; and optional device-side output for ETKDG and forcefield optimization. Blackwell / L-class GPUs (including sm_103/B300) are now supported, the supported RDKit range is now 2025.03.1 through 2026.03.1, and nvMolKit is available via pip install nvmolkit.

Contributors#

  • Kevin Boyd (@scal444)

  • Eva Xue (@evasnow1992)

  • Alireza Moradzadeh (@moradza)

  • Andrei Volgin (@volgin)

Features#

  • GPU-accelerated Torsion Fingerprint Deviation (TFD) for batch all-pairs conformer comparison (#71)

  • GPU-accelerated pairwise conformer RMSD matrix computation by @volgin

  • GPU-accelerated UFF force field, supporting all options that the new BatchedForcefield Python API provides for MMFF (#114)

  • New BatchedForcefield Python API exposing per-molecule control over forcefield minimization (MMFF or UFF), and through it custom MMFF optimization options (max iterations, energy/gradient tolerances, non-bonded cutoff) (#70)

  • Distance and position constraints on forcefield optimization (MMFF and UFF) (#26)

  • Multi-conformer minimization in the BatchedForcefield API

  • HardwareOptions support for MMFF minimization, matching the ETKDG hardware-targeting API

  • Device-side output for ETKDG and forcefield optimization, allowing GPU tensors to flow between nvMolKit calls without round-tripping through host memory (#140)

  • Python autotuning library for the main APIs (nvmolkit.autotune), including ETKDG, forcefield optimization, and substructure search, with configuration serialization (#141)

  • Low-memory fused Butina clustering that computes Tanimoto similarities on the fly with Triton-backed kernels, avoiding the O(N²) distance matrix and enabling clustering of larger fingerprint datasets on a single GPU (#110)

  • Support for Blackwell and L-class GPUs, including sm_103 SASS for B300

Bug Fixes#

  • Fix latent stream-ordering bug in the MMFF/BFGS minimizer that could race with subsequent operations (#172)

  • Fix int32 overflow in substructure pair indexing for batches where numTargets * numQueries exceeds INT32_MAX, which previously caused out-of-bounds writes in hasSubstructMatch and countSubstructMatches (#169)

  • Fix shared-memory overflow in the substructure recursive preprocessor caused by an incorrect config setting (#98)

  • Fix empty result handling in substructure search with uniquify when all matches were already unique (#112)

Miscellaneous#

  • pip wheel distribution pipeline (pip install nvmolkit) with manylinux_2_28 wheels for CPython 3.11-3.14 (#15)

  • RDKit support range is now 2025.03.1 through 2026.03.1

  • Validate batchesPerGpu in HardwareOptions so every consumer gets a clean ValueError instead of a cryptic C++ error from the MMFF / ETKDG layer (#103)

  • Validate neighborlist_max_size in butina() before reaching the GPU (#104)

  • Validate MMFF atom types up front and report every failing molecule instead of hitting a PRECONDITION assertion mid-batch (#106)

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)

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

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)

0.0.1 2025-09-09#

Summary#

Initial release of nvMolKit. Features include:

  • Morgan Fingerprints

  • Tanimoto and Cosine similarity

  • ETKDG conformer generation

  • MMFF optimization