Skip to content

CHANGELOG.md

Latest commit

 

History

History
40 lines (30 loc) · 3.36 KB

CHANGELOG.md

File metadata and controls

40 lines (30 loc) · 3.36 KB
Raw

Changelog

All notable changes to Uni-Dock (the V1 GPU docking engine in unidock/) are documented here.

1.2.0 - 2026-06-11

Hardware support for Blackwell GPUs plus a batch of correctness and crash fixes. The grid-buffer and macrocycle fixes change docked results for the affected inputs (they correct previously-wrong/undefined behavior) — see notes below.

Added

  • Blackwell / CUDA 12.8 support — build for sm_100 (B100/B200/GB200) and sm_120 (RTX 50/B40); CUDA bumped to 12.8 (#184).

Fixed — correctness (results change for affected inputs)

  • Grid buffer sized for the default box — MAX_NUM_OF_GRID_POINT 80³ → 81³ (531,441), so the default 30 Å / 0.375 Å box fits exactly. Previously the bounds check was a release-stripped assert and the grid silently overflowed grid_cuda_t::m_data, corrupting ~20% of results for boxes ≥ 30 Å and crashing (SIGSEGV) on larger boxes (#195, #177; closes #174, #14).
  • Large boxes now error cleanly — boxes exceeding the GPU grid buffer print a clear message and exit instead of silently overflowing / SIGSEGV (#177).
  • Macrocycle (Meeko CG/G) affinities — the ring-closure "glue" energy no longer leaks into the reported affinity (it's a sampling artifact used only to close the ring during optimization). Fixes absurd, non-reproducible scores for macrocycle ligands (#196; closes #102, #192).
  • Out-of-bounds poses are skipped instead of being reported with FLT_MAX (3.4e38) affinities (#182; closes #144).

Fixed — crashes

  • CUDA error 700 (illegal address) in --gpu_batch from a cudaMemset pointer bug (&ig_cuda_gpu → ig_cuda_gpu) (#178; addresses #115/#118/#136).
  • Segfault when a ligand fails to parse in a batch — the bad ligand is now skipped with a warning instead of crashing the batch (#181; closes #112).
  • SIGSEGV on unsupported atom types in SDF (e.g. boron) — SDF atoms are now validated like the PDBQT path and unsupported ones are skipped (#180; closes #138).
  • Intermittent SIGSEGV from a NaN / non-unit axis in angle_to_quaternion — guarded with a renormalize / identity-quaternion fallback (#179; closes #160).
  • CUDA error 700 with flexible residues (--flex) — ≤ 1 flex torsion now docks on GPU; > 1 exits with a clear "supports at most 1" message instead of crashing (#176; closes #159).

Build / CI

  • Fix -DFETCH_BOOST=ON against system Boost ≥ 1.90 — link the fetched Boost::math into the CUDA target so nvcc doesn't pick up the system headers (#189).
  • Repair the (manually-dispatched) benchmark workflow: call the current run_test.py entry point and download the dataset (#194).
  • Pin Python 3.12 for the tools/CI images (openbabel compatibility) (#183, #188); dependency bumps (#185, #186, #187).

Changed / Removed

  • License unified to Apache-2.0 (#153, #155).
  • Removed the built-in receptor processor (it didn't support most user proteins) (#164).

Known limitations

  • Macrocycle sampling: the broken ring isn't held as tightly closed as AutoDock during the GPU search (glue equilibrium ~1.5 Å vs ~0.03 Å), so absolute affinities for macrocycles can be weaker than AutoDock even though scores are now sane and reproducible (tracked in #37).
  • GPU flexible-receptor docking supports at most 1 flex torsion (#175 prototype for more).