CHANGELOG.md

raft 22.10.00 (Date TBD)

Please see https://github.com/rapidsai/raft/releases/tag/v22.10.00a for the latest changes to this development branch.

raft 22.08.00 (17 Aug 2022)

🚨 Breaking Changes

• Update mdspan to account for changes to extents (#751) @divyegala
• Replace csr_adj_graph functions with faster equivalent (#746) @ahendriksen
• Integrate KNN implementation: ivf-flat (#652) @achirkin
• Moving kmeans from cuml to Raft (#605) @lowener

🐛 Bug Fixes

• Relax ivf-flat test recall thresholds (#766) @achirkin
• Restrict the use of ] to CXX 20 only. (#764) @trivialfis
• Update rapids-cmake version for pyraft in update-version.sh (#749) @vyasr

📖 Documentation

• Use documented header template for doxygen (#773) @galipremsagar
• Switch language from None to "en" in docs build (#721) @galipremsagar

🚀 New Features

• Update mdspan to account for changes to extents (#751) @divyegala
• Implement matrix transpose with mdspan. (#739) @trivialfis
• Implement unravel_index for row-major array. (#723) @trivialfis
• Integrate KNN implementation: ivf-flat (#652) @achirkin

🛠️ Improvements

• Use common js and css code (#779) @galipremsagar
• Pin dask & distributed for release (#772) @galipremsagar
• Move cmake to the build section. (#763) @vyasr
• Adding old kmeans impl back in (as kmeans_deprecated) (#761) @cjnolet
• Fix for KMeans raw pointers API (#758) @lowener
• Fix KMeans (#756) @divyegala
• Add inline to nccl_sync_stream() (#750) @seunghwak
• Replace csr_adj_graph functions with faster equivalent (#746) @ahendriksen
• Add wrapper functions for ncclGroupStart() and ncclGroupEnd() (#742) @seunghwak
• Fix variadic template type check for mdarrays (#741) @hlinsen
• RMAT rectangular graph generator (#738) @teju85
• Update conda recipes to UCX 1.13.0 (#736) @pentschev
• Add warp-aggregated atomic increment (#735) @ahendriksen
• fix logic bug in include_checker.py utility (#734) @grlee77
• Support 32bit and unsigned indices in bruteforce KNN (#730) @achirkin
• Ability to use ccache to speedup local builds (#729) @teju85
• Pin max version of cuda-python to 11.7.0 (#728) @Ethyling
• Always add raft::raft_nn_lib and raft::raft_distance_lib aliases (#727) @trxcllnt
• Add several type aliases and helpers for creating mdarrays (#726) @achirkin
• fix nans in naive kl divergence kernel introduced by div by 0. (#724) @mdoijade
• Use rapids-cmake for cuco (#722) @vyasr
• Update Python classifiers. (#719) @bdice
• Fix sccache (#718) @Ethyling
• Introducing raft::mdspan as an alias (#715) @divyegala
• Update cuco version (#714) @vyasr
• Update conda environment pinnings and update-versions.sh. (#713) @bdice
• Branch 22.08 merge branch 22.06 (#712) @cjnolet
• Testing conda compilers (#705) @cjnolet
• Unpin dask & distributed for development (#704) @galipremsagar
• Avoid shadowing CMAKE_ARGS variable in build.sh (#701) @vyasr
• Use unique ptr in print_device_vector (#695) @lowener
• Add missing Thrust includes (#678) @bdice
• Consolidate C++ conda recipes and add libraft-tests package (#641) @Ethyling
• Moving kmeans from cuml to Raft (#605) @lowener

raft 22.06.00 (7 Jun 2022)

🚨 Breaking Changes

• Rng: removed cyclic dependency creating hard-to-debug compiler errors (#639) @MatthiasKohl
• Allow enabling NVTX markers by downstream projects after install (#610) @achirkin
• Rng: expose host-rng-state in host-only API (#609) @MatthiasKohl

🐛 Bug Fixes

• For fixing the cuGraph test failures with PCG (#690) @vinaydes
• Fix excessive memory used in selection test (#689) @achirkin
• Revert print vector changes because of std::vector<bool> (#681) @lowener
• fix race in fusedL2knn smem read/write by adding a syncwarp (#679) @mdoijade
• gemm: fix parameter C mistakenly set as const (#664) @achirkin
• Fix SelectionTest: allow different indices when keys are equal. (#659) @achirkin
• Revert recent cmake updates (#657) @cjnolet
• Don't install component dependency files in raft-header only mode (#655) @robertmaynard
• Rng: removed cyclic dependency creating hard-to-debug compiler errors (#639) @MatthiasKohl
• Fixing raft compile bug w/ RNG changes (#634) @cjnolet
• Get libcudacxx from cuco (#632) @trxcllnt
• RNG API fixes (#630) @MatthiasKohl
• Fix mdspan accessor mixin offset policy. (#628) @trivialfis
• Branch 22.06 merge 22.04 (#625) @cjnolet
• fix issue in fusedL2knn which happens when rows are multiple of 256 (#604) @mdoijade

🚀 New Features

• Restore changes from #653 and #655 and correct cmake component dependencies (#686) @robertmaynard
• Adding handle and stream to pylibraft (#683) @cjnolet
• Map CMake install components to conda library packages (#653) @robertmaynard
• Rng: expose host-rng-state in host-only API (#609) @MatthiasKohl
• mdspan/mdarray template functions and utilities (#601) @divyegala

🛠️ Improvements

• Change build.sh to find C++ library by default (#697) @vyasr
• Pin dask and distributed for release (#693) @galipremsagar
• Pin dask & distributed for release (#680) @galipremsagar
• Improve logging (#673) @achirkin
• Fix minor errors in CMake configuration (#662) @vyasr
• Pulling mdspan fork (from official rapids repo) into raft to remove dependency (#649) @cjnolet
• Fixing the unit test issue(s) in RAFT (#646) @vinaydes
• Build pyraft with scikit-build (#644) @vyasr
• Some fixes to pairwise distances for cupy integration (#643) @cjnolet
• Require UCX 1.12.1+ (#638) @jakirkham
• Updating raft rng host public API and adding docs (#636) @cjnolet
• Build pylibraft with scikit-build (#633) @vyasr
• Add cuda_lib_dir to library_dirs, allow changing UCX/RMM/Thrust/spdlog locations via envvars in setup.py (#624) @trxcllnt
• Remove perf prints from MST (#623) @divyegala
• Enable components installation using CMake (#621) @Ethyling
• Allow nullptr as input-indices argument of select_k (#618) @achirkin
• Update CMake pinning to allow newer CMake versions (#617) @vyasr
• Unpin dask & distributed for development (#616) @galipremsagar
• Improve performance of select-top-k RADIX implementation (#615) @achirkin
• Moving more prims benchmarks to RAFT (#613) @cjnolet
• Allow enabling NVTX markers by downstream projects after install (#610) @achirkin
• Improve performance of select-top-k WARP_SORT implementation (#606) @achirkin
• Enable building static libs (#602) @trxcllnt
• Update ucx-py version (#596) @ajschmidt8
• Fix merge conflicts (#587) @ajschmidt8
• Making cuco, thrust, and mdspan optional dependencies. (#585) @cjnolet
• Some RBC3D fixes (#530) @cjnolet

raft 22.04.00 (6 Apr 2022)

🚨 Breaking Changes

• Moving some of the remaining linalg prims from cuml (#502) @cjnolet
• Fix badly merged cublas wrappers (#492) @achirkin
• Hiding implementation details for lap, clustering, spectral, and label (#477) @cjnolet
• Adding destructor for std comms and using nccl allreduce for barrier in mpi comms (#473) @cjnolet
• Cleaning up cusparse_wrappers (#441) @cjnolet
• Improvents to RNG (#434) @vinaydes
• Remove RAFT memory management (#400) @viclafargue
• LinAlg impl in detail (#383) @divyegala

🐛 Bug Fixes

• Pin cmake in conda recipe to <3.23 (#600) @dantegd
• Fix make_device_vector_view (#595) @lowener
• Update cuco version. (#592) @vyasr
• Fixing raft headers dir (#574) @cjnolet
• Update update-version.sh (#560) @raydouglass
• find_package(raft) can now be called multiple times safely (#532) @robertmaynard
• Allocate sufficient memory for Hungarian if number of batches > 1 (#531) @ChuckHastings
• Adding lap.hpp back (with deprecation) (#529) @cjnolet
• raft-config is idempotent no matter RAFT_COMPILE_LIBRARIES value (#516) @robertmaynard
• Call initialize() in mpi_comms_t constructor. (#506) @seunghwak
• Improve row-major meanvar kernel via minimizing atomicCAS locks (#489) @achirkin
• Adding destructor for std comms and using nccl allreduce for barrier in mpi comms (#473) @cjnolet

📖 Documentation

• Updating docs for 22.04 (#566) @cjnolet

🚀 New Features

• Add benchmarks (#549) @achirkin
• Unify weighted mean code (#514) @lowener
• single-pass raft::stats::meanvar (#472) @achirkin
• Move random package of cuML to RAFT (#449) @divyegala
• mdspan integration. (#437) @trivialfis
• Interruptible execution (#433) @achirkin
• make raft sources compilable with clang (#424) @MatthiasKohl
• Span implementation. (#399) @trivialfis

🛠️ Improvements

• Adding build script for docs (#589) @cjnolet
• Temporarily disable new ops-bot functionality (#586) @ajschmidt8
• Fix commands to get conda output files (#584) @Ethyling
• Link to cuco and add faiss EXCLUDE_FROM_ALL option (#583) @trxcllnt
• exposing faiss::faiss (#582) @cjnolet
• Pin dask and distributed version (#581) @galipremsagar
• removing exclude_from_all from cuco (#580) @cjnolet
• Adding INSTALL_EXPORT_SET for cuco, rmm, thrust (#579) @cjnolet
• Thrust package name case (#576) @trxcllnt
• Add missing thrust includes to transpose.cuh (#575) @zbjornson
• Use unanchored clang-format version check (#573) @zbjornson
• Fixing accidental removal of thrust target from cmakelists (#571) @cjnolet
• Don't add gtest to build export set or generate a gtest-config.cmake (#565) @trxcllnt
• Set main label by default (#559) @galipremsagar
• Add local conda channel while looking for conda outputs (#558) @Ethyling
• Updated dask and distributed to >=2022.02.1 (#557) @rlratzel
• Upload packages using testing label for nightlies (#556) @Ethyling
• Add .github/ops-bot.yaml config file (#554) @ajschmidt8
• Disabling benchmarks building by default. (#553) @cjnolet
• KNN select-top-k variants (#551) @achirkin
• Adding logger (#550) @cjnolet
• clang-tidy support: improved clang run scripts with latest changes (see cugraph-ops) (#548) @MatthiasKohl
• Pylibraft for pairwise distances (#540) @cjnolet
• mdspan PoC for distance make_blobs (#538) @cjnolet
• Include thrust/sort.h in ball_cover.cuh (#526) @akifcorduk
• Increase parallelism in allgatherv (#525) @seunghwak
• Moving device functions to cuh files and deprecating hpp (#524) @cjnolet
• Use dynamic_extent from stdex. (#523) @trivialfis
• Updating some of the ci check scripts (#522) @cjnolet
• Use shfl_xor in warpReduce for broadcast (#521) @akifcorduk
• Fixing Python conda package and installation (#520) @cjnolet
• Adding instructions to install from conda and build using CPM (#519) @cjnolet
• Implement span storage optimization. (#515) @trivialfis
• RNG test fixes and improvements (#513) @vinaydes
• Moving scores and metrics over to raft::stats (#512) @cjnolet
• Random ball cover in 3d (#510) @cjnolet
• Initializing memory in RBC (#509) @cjnolet
• Adjusting conda packaging to remove duplicate dependencies (#508) @cjnolet
• Moving remaining stats prims from cuml (#507) @cjnolet
• Correcting the namespace (#505) @vinaydes
• Passing stream through commsplit (#503) @cjnolet
• Moving some of the remaining linalg prims from cuml (#502) @cjnolet
• Fixing spectral APIs (#496) @cjnolet
• Fix badly merged cublas wrappers (#492) @achirkin
• Fix integer overflow in distances (#490) @RAMitchell
• Reusing shared libs in gpu ci builds (#487) @cjnolet
• Adding fatbin to shared libs and fixing conda paths in cpu build (#485) @cjnolet
• Add CMake install rule for tests (#483) @ajschmidt8
• Adding cpu ci for conda build (#482) @cjnolet
• iUpdating codeowners to use new raft codeowners (#480) @cjnolet
• Hiding implementation details for lap, clustering, spectral, and label (#477) @cjnolet
• Define PTDS via -D to fix cache misses in sccache (#476) @trxcllnt
• Unpin dask and distributed (#474) @galipremsagar
• Replace ccache with sccache (#471) @ajschmidt8
• More README updates (#467) @cjnolet
• CUBLAS wrappers with switchable host/device pointer mode (#453) @achirkin
• Cleaning up cusparse_wrappers (#441) @cjnolet
• Adding conda packaging for libraft and pyraft (#439) @cjnolet
• Improvents to RNG (#434) @vinaydes
• Hiding implementation details for comms (#409) @cjnolet
• Remove RAFT memory management (#400) @viclafargue
• LinAlg impl in detail (#383) @divyegala

raft 22.02.00 (2 Feb 2022)

🚨 Breaking Changes

• Simplify raft component CMake logic, and allow compilation without FAISS (#428) @robertmaynard
• One cudaStream_t instance per raft::handle_t (#291) @divyegala

🐛 Bug Fixes

• Removing extra logging from faiss mr (#463) @cjnolet
• Pin dask & distributed versions (#455) @galipremsagar
• Replace RMM CUDA Python bindings with those provided by CUDA-Python (#451) @shwina
• Fix comms memory leak (#436) @seunghwak
• Fix C++ doxygen documentation (#426) @achirkin
• Fix clang-format style errors (#425) @achirkin
• Fix using incorrect macro RAFT_CHECK_CUDA in place of RAFT_CUDA_TRY (#415) @achirkin
• Fix CUDA_CHECK_NO_THROW compatibility define (#414) @zbjornson
• Disabling fused l2 knn from bfknn (#407) @cjnolet
• Disabling expanded fused l2 knn to unblock cuml CI (#404) @cjnolet
• Reverting default knn distance to L2Unexpanded for now. (#403) @cjnolet

📖 Documentation

• README and build fixes before release (#459) @cjnolet
• Updates to Python and C++ Docs (#442) @cjnolet

🚀 New Features

• error macros: determining buffer size instead of fixed 2048 chars (#420) @MatthiasKohl
• NVTX range helpers (#416) @achirkin

🛠️ Improvements

• Splitting fused l2 knn specializations (#461) @cjnolet
• Update cuCollection git tag (#447) @seunghwak
• Remove libcudacxx patch needed for nvcc 11.4 (#446) @robertmaynard
• Unpin dask and distributed (#440) @galipremsagar
• Public apis for remainder of matrix and stats (#438) @divyegala
• Fix bug in producer-consumer buffer exchange which occurs in UMAP test on GV100 (#429) @mdoijade
• Simplify raft component CMake logic, and allow compilation without FAISS (#428) @robertmaynard
• Update ucx-py version on release using rvc (#422) @Ethyling
• Disabling fused l2 knn again. Not sure how this got added back. (#421) @cjnolet
• Adding no throw macro variants (#417) @cjnolet
• Remove IncludeCategories from .clang-format (#412) @codereport
• fix nan issues in L2 expanded sqrt KNN distances (#411) @mdoijade
• Consistent renaming of CHECK_CUDA and *_TRY macros (#410) @cjnolet
• Faster matrix-vector-ops (#401) @achirkin
• Adding dev conda environment files. (#397) @cjnolet
• Update to UCX-Py 0.24 (#392) @pentschev
• Branch 21.12 merge 22.02 (#386) @cjnolet
• Hiding implementation details for sparse API (#381) @cjnolet
• Adding distance specializations (#376) @cjnolet
• Use FAISS with RMM (#363) @viclafargue
• Add Fused L2 Expanded KNN kernel (#339) @mdoijade
• Update .clang-format to be consistent with all other RAPIDS repos (#300) @codereport
• One cudaStream_t instance per raft::handle_t (#291) @divyegala

raft 21.12.00 (9 Dec 2021)

🚨 Breaking Changes

• Use 64 bit CuSolver API for Eigen decomposition (#349) @lowener

🐛 Bug Fixes

• Fixing bad host->device copy (#375) @cjnolet
• Fix coalesced access checks in matrix_vector_op (#372) @achirkin
• Port libcudacxx patch from cudf (#370) @dantegd
• Fixing overflow in expanded distances (#365) @cjnolet

📖 Documentation

• Getting doxygen to run (#371) @cjnolet

🛠️ Improvements

• Upgrade clang to 11.1.0 (#394) @galipremsagar
• Fix Changelog Merge Conflicts for branch-21.12 (#390) @ajschmidt8
• Pin max dask & distributed (#388) @galipremsagar
• Removing conflict w/ CUDA_CHECK (#378) @cjnolet
• Update RAFT test directory (#359) @viclafargue
• Update to UCX-Py 0.23 (#358) @pentschev
• Hiding implementation details for random, stats, and matrix (#356) @divyegala
• README updates (#351) @cjnolet
• Use 64 bit CuSolver API for Eigen decomposition (#349) @lowener
• Hiding implementation details for distance primitives (dense + sparse) (#344) @cjnolet
• Unpin dask & distributed in CI (#338) @galipremsagar

raft 21.10.00 (7 Oct 2021)

🚨 Breaking Changes

• Miscellaneous tech debts/cleanups (#286) @viclafargue

🐛 Bug Fixes

• Accounting for rmm::cuda_stream_pool not having a constructor for 0 streams (#329) @divyegala
• Fix wrong lda parameter in gemv (#327) @achirkin
• Fix matrixVectorOp to verify promoted pointer type is still aligned to vectorized load boundary (#325) @viclafargue
• Pin rmm to branch-21.10 and remove warnings from kmeans.hpp (#322) @dantegd
• Temporarily pin RMM while refactor removes deprecated calls (#315) @dantegd
• Fix more warnings (#311) @harrism

📖 Documentation

• Fix build doc (#316) @lowener

🚀 New Features

• Add Hamming, Jensen-Shannon, KL-Divergence, Russell rao and Correlation distance metrics support (#306) @mdoijade

🛠️ Improvements

• Pin max dask and distributed versions to 2021.09.1 (#334) @galipremsagar
• Make sure we keep the rapids-cmake and raft cal version in sync (#331) @robertmaynard
• Add broadcast with const input iterator (#328) @seunghwak
• Fused L2 (unexpanded) kNN kernel for NN <= 64, without using temporary gmem to store intermediate distances (#324) @mdoijade
• Update with rapids cmake new features (#320) @robertmaynard
• Update to UCX-Py 0.22 (#319) @pentschev
• Fix Forward-Merge Conflicts (#318) @ajschmidt8
• Enable CUDA device code warnings as errors (#307) @harrism
• Remove max version pin for dask & distributed on development branch (#303) @galipremsagar
• Warnings are errors (#299) @harrism
• Use the new RAPIDS.cmake to fetch rapids-cmake (#298) @robertmaynard
• ENH Replace gpuci_conda_retry with gpuci_mamba_retry (#295) @dillon-cullinan
• Miscellaneous tech debts/cleanups (#286) @viclafargue
• Random Ball Cover Algorithm for 2D Haversine/Euclidean (#213) @cjnolet

raft 21.08.00 (4 Aug 2021)

🚨 Breaking Changes

• expose epsilon parameter to allow precision to to be specified (#275) @ChuckHastings

🐛 Bug Fixes

• Fix support for different input and output types in linalg::reduce (#296) @Nyrio
• Const raft handle in sparse bfknn (#280) @cjnolet
• Add cuco::cuco to list of linked libraries (#279) @trxcllnt
• Use nested include in destination of install headers to avoid docker permission issues (#263) @dantegd
• Update UCX-Py version to 0.21 (#255) @pentschev
• Fix mst knn test build failure due to RMM device_buffer change (#253) @mdoijade

🚀 New Features

• Add chebyshev, canberra, minkowksi and hellinger distance metrics (#276) @mdoijade
• Move FAISS ANN wrappers to RAFT (#265) @cjnolet
• Remaining sparse semiring distances (#261) @cjnolet
• removing divye from codeowners (#257) @divyegala

🛠️ Improvements

• Pinning cuco to a specific commit hash for release (#304) @rlratzel
• Pin max dask & distributed versions (#301) @galipremsagar
• Overlap epilog compute with ldg of next grid stride in pairwise distance & fusedL2NN kernels (#292) @mdoijade
• Always add faiss library alias if it's missing (#287) @trxcllnt
• Use NVIDIA/cuCollections repo again (#284) @trxcllnt
• Use the 21.08 branch of rapids-cmake as rmm requires it (#278) @robertmaynard
• expose epsilon parameter to allow precision to to be specified (#275) @ChuckHastings
• Fix 21.08 forward-merge conflicts (#274) @ajschmidt8
• Add lds and sts inline ptx instructions to force vector instruction generation (#273) @mdoijade
• Move ANN to RAFT (additional updates) (#270) @cjnolet
• Sparse semirings cleanup + hash table & batching strategies (#269) @divyegala
• Revert "pin dask versions in CI (#260)" (#264" (#264)) @ajschmidt8
• Pass stream to device_scalar::value() calls. (#259) @harrism
• Update get_rmm.cmake to better support CalVer (#258) @harrism
• Add Grid stride pairwise dist and fused L2 NN kernels (#250) @mdoijade
• Fix merge conflicts (#236) @ajschmidt8

raft 21.06.00 (9 Jun 2021)

🐛 Bug Fixes

• Update UCX-Py version to 0.20 (#254) @pentschev
• cuco git tag update (again) (#248) @seunghwak
• Revert PR #232 for 21.06 release (#246) @dantegd
• Python comms to hold onto server endpoints (#241) @cjnolet
• Fix Thrust 1.12 compile errors (#231) @trxcllnt
• Make sure we use CalVer when checking out rapids-cmake (#230) @robertmaynard
• Loss of Precision in MST weight alteration (#223) @divyegala

🛠️ Improvements

• cuco git tag update (#243) @seunghwak
• Update CHANGELOG.md links for calver (#233) @ajschmidt8
• Add Grid stride pairwise dist and fused L2 NN kernels (#232) @mdoijade
• Updates to enable HDBSCAN (#208) @cjnolet

raft 0.19.0 (21 Apr 2021)

🐛 Bug Fixes

• Exposing spectral random seed property (#193) @cjnolet
• Fix pointer arithmetic in spmv smem kernel (#183) @lowener
• Modify default value for rowMajorIndex and rowMajorQuery in bf-knn (#173) @viclafargue
• Remove setCudaMallocWarning() call for libfaiss[@v1.7.0 (#167) @trxcllnt](https://github.com/v1.7.0 (#167) @trxcllnt)
• Add const to KNN handle (#157) @hlinsen

🚀 New Features

• Moving optimized L2 1-nearest neighbors implementation from cuml (#158) @cjnolet

🛠️ Improvements

• Fixing codeowners (#194) @cjnolet
• Adjust Hellinger pairwise distance to vaoid NaNs (#189) @lowener
• Add column major input support in contractions_nt kernels with new kernel policy for it (#188) @mdoijade
• Dice formula correction (#186) @lowener
• Scaling knn graph fix connectivities algorithm (#181) @cjnolet
• Fixing RAFT CI & a few small updates for SLHC Python wrapper (#178) @cjnolet
• Add Precomputed to the DistanceType enum (for cuML DBSCAN) (#177) @Nyrio
• Enable matrix::copyRows for row major input (#176) @tfeher
• Add Dice distance to distancetype enum (#174) @lowener
• Porting over recent updates to distance prim from cuml (#172) @cjnolet
• Update KNN (#171) @viclafargue
• Adding translations parameter to brute_force_knn (#170) @viclafargue
• Update Changelog Link (#169) @ajschmidt8
• Map operation (#168) @viclafargue
• Updating sparse prims based on recent changes (#166) @cjnolet
• Prepare Changelog for Automation (#164) @ajschmidt8
• Update 0.18 changelog entry (#163) @ajschmidt8
• MST symmetric/non-symmetric output for SLHC (#162) @divyegala
• Pass pre-computed colors to MST (#154) @divyegala
• Streams upgrade in RAFT handle (RMM backend + create handle from parent's pool) (#148) @afender
• Merge branch-0.18 into 0.19 (#146) @dantegd
• Add device_send, device_recv, device_sendrecv, device_multicast_sendrecv (#144) @seunghwak
• Adding SLHC prims. (#140) @cjnolet
• Moving cuml sparse prims to raft (#139) @cjnolet

raft 0.18.0 (24 Feb 2021)

Breaking Changes 🚨

• Make NCCL root initialization configurable. (#120) @drobison00

Bug Fixes 🐛

• Add idx_t template parameter to matrix helper routines (#131) @tfeher
• Eliminate CUDA 10.2 as valid for large svd solving (#129) @wphicks
• Update check to allow svd solver on CUDA>=10.2 (#125) @wphicks
• Updating gpu build.sh and debugging threads CI issue (#123) @dantegd

New Features 🚀

• Adding additional distances (#116) @cjnolet

Improvements 🛠️

• Update stale GHA with exemptions & new labels (#152) @mike-wendt
• Add GHA to mark issues/prs as stale/rotten (#150) @Ethyling
• Prepare Changelog for Automation (#135) @ajschmidt8
• Adding Jensen-Shannon and BrayCurtis to DistanceType for Nearest Neighbors (#132) @lowener
• Add brute force KNN (#126) @hlinsen
• Make NCCL root initialization configurable. (#120) @drobison00
• Auto-label PRs based on their content (#117) @jolorunyomi
• Add gather & gatherv to raft::comms::comms_t (#114) @seunghwak
• Adding canberra and chebyshev to distance types (#99) @cjnolet
• Gpuciscripts clean and update (#92) @msadang

RAFT 0.17.0 (10 Dec 2020)

New Features

• PR #65: Adding cuml prims that break circular dependency between cuml and cumlprims projects
• PR #101: MST core solver
• PR #93: Incorporate Date/Nagi implementation of Hungarian Algorithm
• PR #94: Allow generic reductions for the map then reduce op
• PR #95: Cholesky rank one update prim

Improvements

• PR #108: Remove unused old-gpubuild.sh
• PR #73: Move DistanceType enum from cuML to RAFT
• pr #92: Cleanup gpuCI scripts
• PR #98: Adding InnerProduct to DistanceType
• PR #103: Epsilon parameter for Cholesky rank one update
• PR #100: Add divyegala as codeowner
• PR #111: Cleanup gpuCI scripts
• PR #120: Update NCCL init process to support root node placement.

Bug Fixes

• PR #106: Specify dependency branches to avoid pip resolver failure
• PR #77: Fixing CUB include for CUDA < 11
• PR #86: Missing headers for newly moved prims
• PR #102: Check alignment before binaryOp dispatch
• PR #104: Fix update-version.sh
• PR #109: Fixing Incorrect Deallocation Size and Count Bugs

RAFT 0.16.0 (Date TBD)

New Features

• PR #63: Adding MPI comms implementation
• PR #70: Adding CUB to RAFT cmake

Improvements

• PR #59: Adding csrgemm2 to cusparse_wrappers.h
• PR #61: Add cusparsecsr2dense to cusparse_wrappers.h
• PR #62: Adding get_device_allocator to handle.pxd
• PR #67: Remove dependence on run-time type info

Bug Fixes

• PR #56: Fix compiler warnings.
• PR #64: Remove cublas_try from cusolver_wrappers.h
• PR #66: Fixing typo get_stream to getStream in handle.pyx
• PR #68: Change the type of recvcounts & displs in allgatherv from size_t[] to size_t* and int[] to size_t*, respectively.
• PR #69: Updates for RMM being header only
• PR #74: Fix std_comms::comm_split bug
• PR #79: remove debug print statements
• PR #81: temporarily expose internal NCCL communicator

RAFT 0.15.0 (Date TBD)

New Features

• PR #12: Spectral clustering.
• PR #7: Migrating cuml comms -> raft comms_t
• PR #18: Adding commsplit to cuml communicator
• PR #15: add exception based error handling macros
• PR #29: Add ceildiv functionality
• PR #44: Add get_subcomm and set_subcomm to handle_t

Improvements

• PR #13: Add RMM_INCLUDE and RMM_LIBRARY options to allow linking to non-conda RMM
• PR #22: Preserve order in comms workers for rank initialization
• PR #38: Remove #include <cudar_utils.h> from raft/mr/
• PR #39: Adding a virtual destructor to raft::handle_t and raft::comms::comms_t
• PR #37: Clean-up CUDA related utilities
• PR #41: Upgrade to cusparseSpMV(), alg selection, and rectangular matrices.
• PR #45: Add Ampere target to cuda11 cmake
• PR #47: Use gtest conda package in CMake/build.sh by default

Bug Fixes

• PR #17: Make destructor inline to avoid redeclaration error
• PR #25: Fix bug in handle_t::get_internal_streams
• PR #26: Fix bug in RAFT_EXPECTS (add parentheses surrounding cond)
• PR #34: Fix issue with incorrect docker image being used in local build script
• PR #35: Remove #include <nccl.h> from raft/error.hpp
• PR #40: Preemptively fixed future CUDA 11 related errors.
• PR #43: Fixed CUDA version selection mechanism for SpMV.
• PR #46: Fix for cpp file extension issue (nvcc-enforced).
• PR #48: Fix gtest target names in cmake build gtest option.
• PR #49: Skip raft comms test if raft module doesn't exist

RAFT 0.14.0 (Date TBD)

New Features

• Initial RAFT version
• PR #3: defining raft::handle_t, device_buffer, host_buffer, allocator classes

Bug Fixes

• PR #5: Small build.sh fixes