certex 0.1.0

The first public release.

Added

• certex.search(xb, xq, k) and certex.Index: exact k-nearest-neighbor search on NVIDIA GPUs by certified
  exclusion, with squared Euclidean, cosine, and inner-product metrics; insertions and removals while searches run;
  filtered search by labels, cuVS bitset filters, or per-query masks; full-precision vectors in GPU or pinned host
  memory; saving and loading.
• Partition thresholds and per-partition local projections, multi-GPU sharding with a shared threshold
  (certex.ShardedIndex), a calibrated cost model that chooses the screening width (certex.tune), and compact
  int8 screening rows.
• k up to 2048.
• Sparse attention: certex.topk (instead of torch.topk(q @ keys.mT, k)), certex.KeyCache for keys that grow
  during decoding, and certex.flashinfer with the top-k transforms of FlashInfer.
• Adapters with the calls of cuVS brute force (certex.cuvs) and Faiss flat indexes (certex.faiss).
• certex.verify(xb, xq, I): checks the neighbors of any library against exhaustive search in double precision.
• The CUDA C++ library as a cuVS module (cuvs::neighbors::certex) and its C API (cuvsCertex*), with Hopper and
  Ampere kernels.
• Examples for every interface: Python search and index, cuVS and Faiss adapters, PyTorch and FlashInfer sparse
  attention, and complete C++ and C programs (examples/cpp, examples/c, built with the C++ library); documentation.
• Documentation that cites the Zarc paper (doi:10.5281/zenodo.23674583),
  which describes the library, beside the CertEx paper of its certified screening.
• tests/test_dropin.py: the same calls on cuVS brute force, Faiss flat indexes, and FlashInfer's
  top_k_page_table_transform and on the certex adapters give the same neighbors.

Install

pip install torch --index-url https://download.pytorch.org/whl/cu129   # PyTorch for CUDA 12.9, as cuVS needs
pip install certex-0.1.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl   # the wheel of your Python

Papers

• Zarc: Exact Nearest Neighbors over a Billion Vectors in Seconds, https://doi.org/10.5281/zenodo.23674583
• CertEx: Exact Visual Retrieval at Scale by Certified Exclusion, https://arxiv.org/abs/2610.21325