v26.10.00

What's Changed

🚨 Breaking Changes

• Apply ABI Breaking Changes for KMeans by @tarang-jain in #2147
• Containerize standalone tarball build by @jameslamb in #2595
• Migrate stream APIs from rmm::cuda_stream_view to cuda::stream_ref by @bdice in #2521
• Iterative CAGRA-Q by @irina-resh-nvda in #1810
• FIX IVF-RabitQ: allow searching after building without serialize round-trip by @achirkin in #2664

🐛 Bug Fixes

• Fix missing trailing barrier in ivf_rabitq blockReduceSum by @hhy3 in #2322
• Use main branch for C ABI baseline by @jrbourbeau in #2388
• Improve C ABI struct change error handling by @jrbourbeau in #2386
• [WIP] Preload cudart before cuvs native library load to prevent downstream class load errors by @rahulgoswami in #2376
• Fix benchmark regressions from #1846 by @achirkin in #2384
• fix knn_merge_parts for inner product by @qwertyforce in #2426
• Fix cuvs-bench opensearch recall and save results by @jrbourbeau in #2369
• Fix CAGRA-ACE regressions from #1846 by @julianmi in #2416
• Remove unused extend code in cagra by @aamijar in #2436
• Fix batched silhouette score reduction aliasing by @csadorf in #2422
• Align views with partial batch extents for SNMG ANN direct merge by @viclafargue in #2441
• Fix devcontainer cache version updates by @bdice in #2479
• Fix silent brute force fallback at aligned vector dimensions by @imotov in #2483
• Fix intermittent ThreadLeak failures in the concurrent IT tests by @imotov in #2499
• Fix cuvs build due to rmm and cuCollections api changes by @jcrist in #2504
• [CI] Update Convergence Criteria to Fix Flaky KMeans Test by @tarang-jain in #2256
• Keep cufile.log out of the Java source tree by @imotov in #2553
• Match merged vectors by document id, not by ordinal position by @imotov in #2556
• Fix hnswlib cmake download only by @aamijar in #2638
• Fix creation of hnswlib target by @KyleFromNVIDIA in #2643
• Lucene: add exact search fallback for very selective filters by @imotov in #2612
• Lucene: Fix GPU search failing on segments with no live vectors by @imotov in #2637
• Disable AnnNNDescentTestUI8_U32 until release/26.10 is complete by @aamijar in #2656

📖 Documentation

• Fix the build_knn_graph sample code by @enp1s0 in #2339
• Adding new docs link(s) to readme. Adding landing page for API refs by @cjnolet in #2228
• Keep the Fern API reference in sync with the sources by @imotov in #2547
• Sync up API docs with code by @imotov in #2561
• Adds a Lucene API section to the Fern docs, generated from java/cuvs-… by @imotov in #2580
• Move Lucene getting-started guide to Fern, expand install and examples by @imotov in #2640

🚀 New Features

• rust: expose cuvsVersionGet via safe Rust by @yan-zaretskiy in #2409
• Fast CAGRA Index Merge by @landrumb in #2352
• Add missing HNSW bindings by @julianmi in #2307
• Generate cubins from cutile at build time and embed them in C++ DSO by @divyegala in #2548
• New 1-nn primitive using cutile by @divyegala in #2552
• Add HNSW Layered Index Support by @julianmi in #2148
• Roaring Bitmap Filter in CAGRA by @divyegala in #2446

🛠️ Improvements

• java: get nightly test matrix from shared-workflows by @jameslamb in #2340
• Update RAPIDS.cmake to log source of rapids-cmake by @arhag23 in #2302
• perf(brute_force): skip csr_to_coo on inner-product filtered search by @maxwbuckley in #2128
• Remove libclang pin from clang dependency group by @vyasr in #2263
• enforce 'yamllint' checks by @jameslamb in #2357
• Main merge release/26.08 by @cjnolet in #2392
• Brings the Java build in line with other clients. by @imotov in #2361
• Fix forward merge from 26.08 into main by @cjnolet in #2411
• ensure nightly builds always produce new packages, expand 'changed-files' lists by @jameslamb in #2414
• Factor CAGRA graph kernels into shared translation unit by @landrumb in #2412
• perf(brute_force): add scatter-gather path and fix filtered-search dispatch thresholds by @maxwbuckley in #2321
• Add cuvs-bench-elastic: HTTP backend for Elasticsearch GPU vector search by @afourniernv in #1907
• Add Gaussian Mixture Models (GMM) by @Intron7 in #2248
• fix(update-version): don't blanket replace the word main with release branch in README.md by @gforsyth in #2421
• wheels: enforce 'abi3audit' checks by @jameslamb in #2439
• Cuvs lucene move by @imotov in #2353
• Use librtcx for JIT-LTO by @arhag23 in #2311
• ci: avoid triggered-by-schedule and triggered-by-merge builds cancelling each other by @jameslamb in #2486
• Make run_spotless.sh more verbose and strict by @imotov in #2485
• update_dataset CAGRA by @aamijar in #2427
• Improve the exception message when cuvs-java is not supported by @imotov in #2495
• Add cuVS overview image to docs home page by @cjnolet in #2514
• ACE: Harden npartitions Check by @julianmi in #2282
• CAGRA-ACE: Make Disk Workspaces Transactional and Non-destructive by @julianmi in #2336
• Replace O(N*degree^2) dedup with per-thread bit-packed bitset by @jamxia155 in #2437
• Lucene: preserve throwable types in accelerated vector formats by @shaunakkapur in #2515
• test: isolate library wheel smoke test, add missing 'cuda-pathfinder' dependency for 'libcuvs' by @jameslamb in #2524
• Delete stale Panama bindings instead of generating on top of them by @imotov in #2511
• Add KvikIO to Improve I/O Throughput by @julianmi in #2257
• Dry run: preprocessing::quantize::scalar by @achirkin in #2519
• Fix: reduce recall calculation time at small batch by @sherylll in #2513
• Vamana build optimization set and fp16 support by @bkarsin in #2264
• Enable security scanning via NVIDIA/security-workflows suite by @gmanal in #2430
• Lucene: Merge CAGRA indexes on the GPU instead of rebuilding from vectors by @imotov in #2497
• Adopt CUDA stream compatibility accessors by @bdice in #2557
• Move Java example projects under examples/java/ by @imotov in #2543
• Optimize cuvs-lucene CAGRA_HNSW index build; add example by @jamxia155 in #2481
• Fix CMake discovery for wheel dependencies by @bdice in #2565
• wheels: get 'tileiras' from system install, not 'cuda-toolkit' wheels by @jameslamb in #2564
• [ANN_BENCH] Remove Stream Pool Creation from Individual Wrappers by @tarang-jain in #2526
• Revert "Optimize cuvs-lucene CAGRA_HNSW index build; add example (#2481)" by @imotov in #2597
• Enable ARM (aarch64) support for cuvs-java and cuvs-lucene by @imotov in #2541
• Optimize transfer/compute overlap in out-of-core KMeans by @viclafargue in #2538
• Add support for BBQ compressed datasets in NN-Descent and CAGRA by @lowener in #2654

New Contributors

• @hhy3 made their first contribution in #2322
• @landrumb made their first contribution in #2412
• @rahulgoswami made their first contribution in #2376
• @afourniernv made their first contribution in #1907
• @sherylll made their first contribution in #2513
• @gmanal made their first contribution in #2430

Full Changelog: v26.10.00a...v26.10.00