v26.04.00

What's Changed

🚨 Breaking Changes

• Use HNSW GPU Hierarchy by Default by @julianmi in #1617
• Backport "Default to static linking of libcudart" by @KyleFromNVIDIA in #1918
• enforce a floor on libnvjitlink, build wheels with CUDA 13.0.x, test wheels against mix of CTK versions by @jameslamb in #1862

🐛 Bug Fixes

• Add float16 support for CAGRA extend by @lowener in #1620
• Fix C examples CUDAToolkit dependency by @tfeher in #1626
• clang-tidy: remove AnalyzeTemporaryDtors key by @achirkin in #1778
• Fix build.sh: build cuvs_c for examples and fix --gpu-arch parsing by @achirkin in #1779
• Better handling of batching of search in MG replicated mode by @viclafargue in #1718
• Add nvjitlink to cuda-toolkit pip extras for cusparse compatibility by @bdice in #1794
• [REVIEW] Fix: for balanced kmeans use grid.x for adjust_centers to avoid grid.y overflow for >262K centroids by @Nischal1729 in #1805
• Fix thrust header by @aamijar in #1817
• Fix persistent CAGRA regressions by @achirkin in #1800
• Fix thrust header by @aamijar in #1825
• Set cudaFuncAttributeMaxDynamicSharedMemorySize with thread-safety by @mythrocks in #1771
• [REVIEW] Move from thrust::make_counting_iterator to cuda::make_counting_iterator by @mythrocks in #1826
• Fixed cuvs benchmark debug build issue (linker step fail) by @irina-resh-nvda in #1599
• Graph degree equals intermediate graph degree bug fix by @irina-resh-nvda in #1834
• FAISS patch for thrust_counting_iterator.h by @aamijar in #1844
• Use 1D grid calculations in epsilon_neighborhood by @divyegala in #1847
• Fix setting CAGRA graph build algo to iterative search by default by @achirkin in #1864
• Faiss suppress warning 611 by @aamijar in #1879
• Pin faiss to 1.14.0 by @aamijar in #1885
• Make some minor fixes to JIT+LTO functionality by @KyleFromNVIDIA in #1896
• Revert "Default to static linking of libcudart" by @KyleFromNVIDIA in #1925
• Fix hanging pytests by @aamijar in #1924
• Disallow programmatic stream serialization in JIT kernel launches by @divyegala in #1932
• IVF-Flat: fix irrelevant assert in the fused kernel mode by @achirkin in #1941
• [REVIEW] cuVS bench: Fix cudaFuncSetAttribute not being called when CAGRA search switches kernel variants by @irina-resh-nvda in #1851
• Pin openblas for aarch64 by @tarang-jain in #1965
• Null JIT kernel launch config by @divyegala in #1974
• [BUG] Fix Vamana Serialization by @tarang-jain in #1966

📖 Documentation

• remove docs references to cuvs-bench-datasets, install a C compiler in pre-commit env by @jameslamb in #1736
• Adding simple tarball install to build and install docs by @cjnolet in #1868
• Fixing nccl link in tarball install instructions by @cjnolet in #1887
• Add developer and user guides for JIT by @divyegala in #1876
• PCA docs by @aamijar in #1949
• Fix a few typos and .rst link syntax by @jrbourbeau in #1973
• Update docs footer year by @aamijar in #1958
• Doc improvements by @aamijar in #1978

🚀 New Features

• JIT compile interleaved_scan_kernel for CUDA 13 by @divyegala in #1405
• [REVIEW] L1 distance support for iterative search CAGRA build by @yan-zaretskiy in #1831
• [FEA] Inertia Computation for Balanced KMeans and Add Option for Weighted Inertia by @tarang-jain in #1880
• PCA preprocessor by @aamijar in #1808
• [FEA] Add Batching to KMeans by @tarang-jain in #1886

🛠️ Improvements

• Automatic Partition Count Derivation for ACE by @julianmi in #1603
• Add filter for cagra::merge by @benfred in #1496
• Ivf_flat extends golang APIs by @cpegeric in #1600
• Drop Python 3.10 support by @gforsyth in #1748
• tighten wheel size limits, expand CI-skipping logic, other small build changes by @jameslamb in #1751
• Migrate hash strategy to use the new cuco::static_map by @PointKernel in #1462
• Update raft headers by @aamijar in #1763
• remove pip.conf migration code in CI scripts, update CI-skipping rules by @jameslamb in #1760
• Convert non-type template parameters to runtime parameters in CAGRA search to cut binary size by @seunghwak in #1498
• Rename lanczos by @aamijar in #1759
• CI: build with CUDA 13.1.1 by @jameslamb in #1766
• Fixes for stricter compilers by @maxwbuckley in #1703
• fix cpu_search call by including k argument by @benfred in #1785
• Use GHA id-token for sccache-dist auth token by @trxcllnt in #1790
• Remove cagra_optimize.hpp by @aamijar in #1791
• Remove unused use_norms constant from kernel_sm60.cuh by @maxwbuckley in #1705
• Remove unused variable read_idx from query loop by @maxwbuckley in #1706
• remove gitutils by @jameslamb in #1797
• refactor: build wheels and conda packages using Python limited API by @gforsyth in #1788
• Add pluggable backend architecture to cuvs-bench by @jnke2016 in #1536
• Use Specific CCCL Includes by @divyegala in #1806
• Replace thrust::tuple with cuda::std::tuple by @miscco in #1811
• check-nightly-ci: update to new version by @jameslamb in #1813
• check-nightly-ci: remove testing config by @jameslamb in #1824
• Refactor JIT LTO kernel generation by @KyleFromNVIDIA in #1812
• Move test_compute_matrix_product.py to cpp/tests by @KyleFromNVIDIA in #1836
• Drop uses of thrust/functional.h by @miscco in #1835
• Better ncv param spectral embedding edge case by @aamijar in #1828
• Modernize the uses of raft in cuVS by @achirkin in #1837
• refactor(limited api): add explicit wheel.py-api to pyproject.toml
  by @gforsyth in #1852
• Update Cython lower bound pin to 3.2.2 by @vyasr in #1858
• Add support for Python 3.14 by @gforsyth in #1845
• fix(rust): change Index::search to take &self instead of self by @zbennett10 in #1839
• Remove pytest upper bound pin by @vyasr in #1867
• Readme: vecflow paper in references by @aamijar in #1874
• Refactor knn graph build params to remove graph_build_types.hpp by @jinsolp in #1235
• Remove IndexWrapper by @aamijar in #1792
• Use C linkage for JIT LTO kernels by @divyegala in #1909
• Prevent nested parallelism in HNSW bench by @julianmi in #1895
• Add a script to check for breaking C ABI changes by @benfred in #1749
• Improve the IVF-PQ Coarse Batch Size Workspace Estimation by @julianmi in #1937
• ScaNN: Fix AVQ prefetch by @rmaschal in #1899

New Contributors

• @cpegeric made their first contribution in #1600
• @PointKernel made their first contribution in #1462
• @seunghwak made their first contribution in #1498
• @Nischal1729 made their first contribution in #1805
• @jnke2016 made their first contribution in #1536
• @zbennett10 made their first contribution in #1839
• @jrbourbeau made their first contribution in #1973

Full Changelog: v26.04.00a...v26.04.00