Faiss Files

Highlights

• Metal GPU backend expansion — new MetalIndexIVFFlat with IVF scan/merge kernels and expanded top-k support (#5202); Metal now enabled by default on Apple Silicon machines (#5280)
• TurboQuant in ScalarQuantizer — full Algorithm 2 (QJL stage) with SIMD and optimizations (#5170)
• Sapphire Rapids (SPR) optimizations — ScalarQuantizer L2/IP (#5173), VPOPCNTDQ-based HammingComputer (#5183), and VPOPCNTDQ-based RaBitQ kernel (#5149)
• faiss-gpu pip wheels — new GPU wheel packaging (#5131); musllinux wheels re-enabled for faiss-cpu (#5299)
• SVS static vamana support (#5224)
• HNSW is_similarity mode for IP/similarity metrics (#5246, [#5226])
• HNSW search performance — reserve VisitedTableSet capacity to avoid rehashes (#5290), search_from_candidate_unbounded templatized for VisitedTable devirtualization (#5270), and runtime checks avoided in VisitedTable (#5234)
• NEON FINE_SIZE=2 specializations for Index2LevelDecoderImpl and IndexPQDecoder (#5255)
• cuVS upgraded to 26.06 (#5240); CI updated to ROCm 7 (#5196)
• Deserialization hardening — bool-field validation (#5279) and null inner-index rejection in IDMap / BinaryFromFloat (#5239)
• Build robustness — fix non-AVX2 import SIGILL via .rodata partitioning tables (#5298), Windows ARM64 (MSVC) NEON build fix (#5274), and AVX512_SPR dispatch fix on AMD (#5281)

Full Changelog

Added

• [261d8f] Add IVFSQTurboQSearchParameters to init.pyi stub (#5304)
• [684a32] Add manual Faiss nightly workflow dispatch (#5300)
• [10b6b2] Add MetalIndexIVFFlat with IVF scan/merge kernels and expanded top-k support (#5202)
• [74d361] TurboQuant in ScalarQuantizer: full Algorithm 2 (QJL stage) with SIMD and optimizations (#5170)
• [506600] Add Python HNSW tutorial (#5260)
• [3cff0f] Add Sapphire Rapids optimizations for ScalarQuantizer (L2, IP) (#5173)
• [46ef80] Support IndexIDMap/IndexIDMap2 in reverse_index_factory (#5266)
• [420158] Add VPOPCNTDQ-based HammingComputer for Sapphire Rapids+ (#5183)
• [d8f1c2] Implement NEON-based FINE_SIZE=2 specializations for Index2LevelDecoderImpl and IndexPQDecoder (#5262)
• [0951b5] Support user provided blas library (#5189)
• [215740] SVS static vamana support (#5224)
• [d24ad6] Add is_similarity mode to HNSW (#5246)
• [aa332c] Implement NEON-based FINE_SIZE=2 specializations for Index2LevelDecoderImpl and IndexPQDecoder (#5255)
• [5d9aae] Add faiss-gpu pip wheel packaging (#5131)
• [6f1cf6] Add VPOPCNTDQ-based RaBitQ kernel for Sapphire Rapids+ (#5149)

Changed

• [262fc3] Re-enable musllinux wheels for faiss-cpu (#5299)
• [1cdc37] Run CI on push to main to refresh ccache cache (#5291)
• [379ee7] Reserve VisitedTableSet capacity to avoid rehashes during HNSW search (#5290)
• [6513a2] Enable Metal by default on Apple machines (#5280)
• [fe46c3] Validate bool fields during deserialization (#5279)
• [480f91] Type imbalance_factor and wire the .pyi stub into the buck build (#5269)
• [e60bae] Templatize search_from_candidate_unbounded for VisitedTable devirtualization (#5270)
• [c00019] Accelerate ScalarQuantizer::QT_bf16 with AVX512-BF16. (#4889)
• [7504fc] Upgrade CUVS Version to 26.06 (#5240)
• [d12683] facebook-unused-include-check in IndexBinaryIVF.cpp (#5263)
• [99d901] facebook-unused-include-check in distances_simd.cpp (#5264)
• [2f0368] facebook-unused-include-check in hamming_avx2.cpp (#5265)
• [0c7275] Remove unused include of platform_macros.h in partitioning.cpp
• [e6b8f6] IndexHNSW: use HNSW::is_similarity for IP/similarity metrics + tests (#5226)
• [c0575f] avoid runtime checks in VisitedTable (#5234)
• [1cb760] Eliminate per-code denormalization in uniform SQ distance computation (#5166)
• [f29d86] Revert D106693266: Implement NEON-based FINE_SIZE=2 specializations for Index2LevelDecoderImpl and IndexPQDecoder
• [ef96e3] Updating CI to ROCm 7 (#5196)
• [f5217d] facebook-unused-include-check in IndexBinaryHNSW.cpp (#5251)
• [3e6ed9] facebook-unused-include-check in IndexBinaryIVF.cpp (#5252)
• [108868] Reject null inner index in IDMap and BinaryFromFloat deserialization (#5239)
• [a64b54] Add IndexLattice r2 limit to cap decode-cache build cost (#5238)
• [099371] faiss: Replace remaining get_single_code calls with ScopedCodes (#5248)
• [910e43] facebook-unused-include-check in hamming_avx2.cpp (#5242)
• [d581f2] Use per-SIMD TU scan for standalone PQ (AVX2 gather inlining) (#5233)

Fixed

• [20afed] make intentional cudaGetLastError() error-clears explicit ((void)) for clang21 -- fixes S674096 (#5302)
• [e420e9] Move partitioning shifts tables to .rodata to fix non-AVX2 import SIGILL (#5298)
• [15bd82] Fix cuVS nightly (#5273)
• [17cc96] Fix AVX512_SPR dispatch on AMD: require AVX512_FP16 CPUID (#5281)
• [e69bfe] Stabilize RaBitQ tests on AVX512_SPR by switching to cross-level equivalence (#5277)
• [c4c651] Stub fixes: knn torch overload, ResidualCoarseQuantizer ctor, remove duplicate I/O defs (#5283)
• [eb4c1e] Fix Windows ARM64 (MSVC) build broken by NEON SIMD templatization (#5274)
• [ab6323] Install missing InvertedListScannerStats.h header (#5268)
• [140541] Fix broken fbcode//faiss/tests:test_index_binary - test_replicas (test_index_binary.TestReplicasAndS (#5258)
• [34eb98] fix(python): int64 coercions for MapLong2Long + InvertedLists DOWNCAST chain (#5257)
• [d492af] faiss: initialize ids_tab to -1 in Top1BlockResultHandler::begin_multiple (#5249)
• [a4e417] Fix: static SIMD dispatch falls to scalar for avx512_spr/avx512/arm_sve builds (#5057)
• [3edc6e] Guard Panorama autovec pragmas against nvcc frontend (#5241)
• [c7689c] Fix ODR violation in ScannerMixIn by adding SIMDLevel template parameter (#5148)
• [ca87f4] Open SIFT demo data in binary mode (#5213)

Full diff: https://github.com/facebookresearch/faiss/compare/v1.14.2...v1.14.3