smsd-cuda

smsd-cuda provides optional NVIDIA CUDA screening on Linux x86_64 and needs the same version of smsd:

python -m pip install "smsd==8.0.2" "smsd-cuda==8.0.2"

The GPU computes integer overlaps; scores use the CPU expressions and hits keep target order. Matching runs on the CPU. The backend keeps prepared corpus fingerprints on the device for reuse. ECFP4 Tanimoto screening and the atom-count MCS similarity bound are distinct operations.

import smsd

corpus = smsd.TargetCorpus.from_smiles(["CCO", "CCN", "CCC"])
print(smsd.gpu_device_info())
corpus.prewarm(num_threads=1)
assert corpus.screen_many(["CCO", "CCN"], 1.0, num_threads=1) == [[0], [1]]
assert corpus.screen_targets_many(["CCO", "CCN"], 1.0, num_threads=1) == [[0], [1]]

smsd finds the package at run time. Automatic ECFP4 screening uses CUDA on a Tesla T4 with compute capability 7.5 when the prepared corpus has at least 100,000 targets. Smaller corpora and other devices use the CPU. Label-bound and substructure candidate screening use the CPU. gpu_is_available() and gpu_device_info() report backend availability, not whether a call used the device. Missing devices, allocation failure and device loss use the CPU. SMSD_DISABLE_GPU=1 disables the GPU at run time.

Corpus substructure searches can reject impossible molecule candidates with a conservative element-path screen before CPU matching. SMARTS queries, untyped queries and finite node limits bypass this screen. Target results keep their original slots, atom numbering and search status.

Use screen_targets with an explicit threshold before MCS when an atom-count Tanimoto bound is wanted; this bound does not represent bond Tanimoto or weighted MCS scores. ECFP similarity is not a substructure test.

Requirements

See the supported-platforms table for Python versions, operating-system requirements and GPU Docker support. The nvidia-cuda-runtime-cu12 dependency supplies the CUDA runtime; the host needs a compatible NVIDIA driver. The CUDA toolkit is not needed to install the wheel.

Building

Build from an SMSD checkout on Linux x86_64 with the CUDA 12.8 toolkit:

python -m pip wheel packaging/smsd-cuda --no-deps -w dist

Its version must equal the smsd version of the checkout; the build stops otherwise. Building requires nvcc; the smsd package itself never needs CUDA.

Licence

Apache License 2.0; see LICENSE and NOTICE. The package bundles no NVIDIA library.