v26.02.00

What's Changed

🚨 Breaking Changes

• Add thread pool as function parameters to C++ API by @kingcrimsontianyu in #876
• Default to static linking of libcudart by @bdice in #885

🚀 New Features

• Expose cuFile stream register API to Python by @kingcrimsontianyu in #893

🛠️ Improvements

• Use strict priority in CI conda tests by @bdice in #877
• Use strict priority in CI conda tests by @bdice in #881
• Remove alpha specs from non-RAPIDS dependencies by @bdice in #883
• Enable merge barriers by @KyleFromNVIDIA in #884
• Add per-block-device thread pools for multi-drive I/O parallelism by @kingcrimsontianyu in #880
• Add devcontainer fallback for C++ test location by @bdice in #887
• Remove deprecated config setters and getters from python API by @kingcrimsontianyu in #882
• Empty commit to trigger a build by @bdice in #889
• Use SPDX license identifiers in pyproject.toml, bump build dependency floors by @jameslamb in #890
• Add CUDA 13.1 support by @bdice in #888
• build and test against CUDA 13.1.0 by @jameslamb in #891
• Empty commit to trigger a build by @jameslamb in #897
• Use main shared-workflows branch by @jameslamb in #900
• wheel builds: react to changes in pip's handling of build constraints by @mmccarty in #902
• fix(build): build package on merge to release/* branch by @gforsyth in #907

New Contributors

• @mmccarty made their first contribution in #902

Full Changelog: v26.02.00a...v26.02.00

v25.12.00

What's Changed

🚨 Breaking Changes

• Remove ability to build kvikio without CUDA by @vyasr in #829
• Minimize transitive includes by moving NVTX-related code to the detail namespace by @kingcrimsontianyu in #851
• Implement simple username-based authentication for WebHDFS by @kingcrimsontianyu in #859
• Add opportunistic Direct I/O support to POSIX backend by @kingcrimsontianyu in #863
• Require CUDA 12.2+ by @jakirkham in #860

🐛 Bug Fixes

• Fix an out-of-bounds memory access in file open flags checking by @kingcrimsontianyu in #853
• Restore isort for Cython files by @KyleFromNVIDIA in #864
• refactored update-version.sh to handle new branching strategy by @rockhowse in #868
• Pin Cython pre-3.2.0 and PyTest pre-9 by @jakirkham in #869

📖 Documentation

• Use current system architecture in conda environment creation command by @bdice in #867

🛠️ Improvements

• Simplify async validation condition by @vyasr in #842
• Update RAPIDS_BRANCH, codify changes in update-version.sh by @KyleFromNVIDIA in #843
• Add named parameters for S3 credentials to Python API by @TomAugspurger in #846
• Create utility class URL builder and URL encoder. Fix a tricky URL path issue by @kingcrimsontianyu in #844
• Make parameter name more consistent for S3 utility functions by @kingcrimsontianyu in #848
• Enable sccache-dist connection pool by @trxcllnt in #852
• Use main in RAPIDS_BRANCH by @bdice in #854
• Use main branch by @bdice in #855
• Use ruff and ruff-format instead of flake8, black, isort by @KyleFromNVIDIA in #856
• Use main shared-workflows branch by @bdice in #857
• Use SPDX for all copyright headers by @KyleFromNVIDIA in #858
• Use RAPIDS_BRANCH in cmake-format invocations that need rapids-cmake configs by @bdice in #862
• Disable benchmark builds by default, enable in CI by @PointKernel in #866
• Use sccache-dist build cluster for conda and wheel builds by @trxcllnt in #865

New Contributors

• @PointKernel made their first contribution in #866
• @rockhowse made their first contribution in #868

Full Changelog: v25.12.00a...v25.12.00

v25.10.00

🚨 Breaking Changes

• Support access to public S3 (#820) @kingcrimsontianyu
• Reduce duplication between compat manager and defaults (#816) @vyasr
• Devendor libnvcomp from libkvikio (#805) @bdice
• Remove Python nvCOMP bindings and Zarr 2 support (#798) @vuule

🐛 Bug Fixes

• Keep string alive until nvtxDomainResourceCreate (#832) @vyasr
• Fix a bug in public S3 inference (#831) @kingcrimsontianyu
• Fix CUDA 13 handling of libcufile on aarch64 (#827) @bdice
• Skip max_device_cache_size setter when BAR1 memory isn't present on the GPUs in the system (#814) @ahoyle-nvidia
• Fix an S3 parsing bug in the open function. Improve regex usage (#810) @kingcrimsontianyu
• Remove remaining nvcomp references (#801) @vyasr
• Revert "Set compiler versions in context (#755)" (#784) @vyasr
• Relax mmap read requirement. Improve error message. (#781) @kingcrimsontianyu

🚀 New Features

• Support access to public S3 (#820) @kingcrimsontianyu
• Upgrade to nvCOMP 5.0.0.6 (#800) @vuule
• Remove Python nvCOMP bindings and Zarr 2 support (#798) @vuule
• Support WebHDFS (2/2): Python binding (#791) @kingcrimsontianyu
• Support file size query for S3 presigned URL (#789) @kingcrimsontianyu
• Support WebHDFS (1/2): C++ implementation (#788) @kingcrimsontianyu
• Support file-backed mapping (2/n): host/device read Python binding (#742) @kingcrimsontianyu

🛠️ Improvements

• Empty commit to trigger a build (#840) @msarahan
• Configure repo for automatic release notes generation (#825) @AyodeAwe
• Use C++20 for KvikIO main library (#819) @kingcrimsontianyu
• Explicitly manage TLS/SSL CA paths for remote I/O (#817) @kingcrimsontianyu
• Reduce duplication between compat manager and defaults (#816) @vyasr
• Added KVIKIO_REMOTE_VERBOSE option (#815) @TomAugspurger
• Use branch-25.10 again (#812) @jameslamb
• Update rapids-dependency-file-generator (#809) @KyleFromNVIDIA
• Add a unified remote I/O interface that infers the endpoint type from URL (2/2): Python binding (#808) @kingcrimsontianyu
• Devendor libnvcomp from libkvikio (#805) @bdice
• Remove more nvcomp packaging for conda (#804) @bdice
• Build and test with CUDA 13.0.0 (#803) @jameslamb
• Optionally require zarr>=3.0.0 (#802) @TomAugspurger
• Use build cluster in devcontainers (#797) @trxcllnt
• Improve KvikIO Python binding performance by releasing GIL wherever deemed necessary (#796) @kingcrimsontianyu
• Update rapids_config to handle user defined branch name (#794) @robertmaynard
• Add a unified remote I/O interface that infers the endpoint type from URL (1/2): C++ implementation (#793) @kingcrimsontianyu
• Update rapids-build-backend to 0.4.0 (#790) @KyleFromNVIDIA
• Fix C++20 warning in the mmap test (#785) @kingcrimsontianyu
• Allow latest OS in devcontainers (#780) @bdice
• Update build infra to support new branching strategy (#776) @robertmaynard
• Use GCC 14 in conda builds. (#756) @vyasr
• Use C++20 standard (#749) @kingcrimsontianyu

[NIGHTLY] v25.12.00

🔗 Links

• Development Branch
• Compare with main branch

🚨 Breaking Changes

• Remove ability to build kvikio without CUDA (#829) @vyasr

🐛 Bug Fixes

• Fix an out-of-bounds memory access in file open flags checking (#853) @kingcrimsontianyu

🛠️ Improvements

• Enable sccache-dist connection pool (#852) @trxcllnt
• Make parameter name more consistent for S3 utility functions (#848) @kingcrimsontianyu
• Add named parameters for S3 credentials to Python API (#846) @TomAugspurger
• Create utility class URL builder and URL encoder. Fix a tricky URL path issue (#844) @kingcrimsontianyu
• Update RAPIDS_BRANCH, codify changes in update-version.sh (#843) @KyleFromNVIDIA
• Simplify async validation condition (#842) @vyasr
• Remove ability to build kvikio without CUDA (#829) @vyasr

v25.08.00

🚨 Breaking Changes

• Remove CUDA 11 from dependencies.yaml (#745) @KyleFromNVIDIA
• stop uploading packages to downloads.rapids.ai (#734) @jameslamb

🐛 Bug Fixes

• Fix remote I/O file size overflow in Python binding (#766) @kingcrimsontianyu
• Fix a case sensitivity bug in S3 remote I/O (#765) @kingcrimsontianyu
• Use RAPIDS' CUDA major-minor version in Conda's cuda_version (#758) @vyasr
• Fix unit test UB resulting from incorrect initialization (#751) @kingcrimsontianyu
• Add an overload getenv_or that supports env var alias. Add new env var KVIKIO_NUM_THREADS. Fix UB. (#735) @kingcrimsontianyu

📖 Documentation

• Remove CUDA 11 mentions from docs (#769) @pentschev
• add docs on CI workflow inputs (#764) @jameslamb

🚀 New Features

• Add the utility function to clear page cache (#741) @kingcrimsontianyu
• Support file-backed mapping (1/n): host/device read C++ implementation (#740) @kingcrimsontianyu

🛠️ Improvements

• Use libnvcomp conda package (#774) @bdice
• fix(docker): use versioned -latest tag for all rapidsai images (#771) @gforsyth
• Remove CUDA 11 workarounds (#770) @pentschev
• Use CUDA 12.9 in Conda, Devcontainers, Spark, GHA, etc. (#762) @jakirkham
• refactor(shellcheck): enable for all files and fix remaining warnings (#760) @gforsyth
• Remove nvidia and dask channels (#759) @vyasr
• Set compiler versions in context (#755) @jakirkham
• Use variants to produce separate builds with and without cufile support (#754) @vyasr
• Only use cufile on x86 (#753) @vyasr
• refactor(rattler): remove explicit libnuma pin now that upstream feedstock is fixed (#752) @gforsyth
• Remove CUDA 11 from dependencies.yaml (#745) @KyleFromNVIDIA
• Add the conda dependency libnuma necessary for GDS I/O (#744) @kingcrimsontianyu
• Remove CUDA 11 devcontainers and update CI scripts (#738) @bdice
• refactor(rattler): remove cuda11 options and general cleanup (#737) @gforsyth
• stop uploading packages to downloads.rapids.ai (#734) @jameslamb
• Forward-merge branch-25.06 into branch-25.08 (#722) @gforsyth
• resolve forward-merge from branch-25.06 to branch-25.08 (#710) @gforsyth

v25.06.00

🚨 Breaking Changes

• Fix the S3 endpoint constructor ambiguity (#700) @kingcrimsontianyu
• Add support for temporary AWS credentials to access S3 (#693) @joosthooz

🐛 Bug Fixes

• [pytest] disable no-data-collected warning (#746) @madsbk
• Revert "Provide API to detect C2C platform and add NVML shim" (#731) @vyasr
• Check whether memory resource/BAR1 are supported (#729) @pentschev
• Use try...finally... to ensure buffer is released (#728) @jakirkham
• Fix benchmark code on MIG (#703) @jakirkham
• Fix the S3 endpoint constructor ambiguity (#700) @kingcrimsontianyu
• zarr-python 3.0.7 compatibility (#699) @TomAugspurger
• Vendor RAPIDS.cmake (#695) @bdice
• Fix setting region in C++ constructor instead of via the env var (#691) @joosthooz

📖 Documentation

• Fix doc build error (#723) @kingcrimsontianyu
• Fix comment typo in buffer_register (#669) @JigaoLuo

🚀 New Features

• Provide API to detect C2C platform and add NVML shim (#721) @kingcrimsontianyu
• Deprecate nvCOMP python API (#714) @kingcrimsontianyu
• Add the feature to check page cache residency information for a file (#704) @kingcrimsontianyu
• Add support for temporary AWS credentials to access S3 (#693) @joosthooz

🛠️ Improvements

• use 'rapids-init-pip' in wheel CI, other CI changes (#725) @jameslamb
• Finish CUDA 12.9 migration and use branch-25.06 workflows (#719) @bdice
• Update to clang 20 (#718) @bdice
• Quote head_rev in conda recipes (#717) @bdice
• Build and test with CUDA 12.9.0 (#715) @bdice
• feat: add support for Python 3.13 (#708) @gforsyth
• refactor(wheels): vendor nvcomp as temporary workaround (#702) @gforsyth
• Download build artifacts from Github for CI (#698) @VenkateshJaya
• Improve unit tests; add a dedicated macro for Linux system calls (#697) @kingcrimsontianyu
• Remove usage of APIs outside the limited API (#692) @vyasr
• Port all conda recipes to rattler-build (#678) @gforsyth
• Make page size portable by querying it at runtime (#677) @kingcrimsontianyu
• chore(conda): enable strict channel priority for conda build (#674) @gforsyth
• Moving wheel builds to specified location and uploading build artifacts to Github (#672) @VenkateshJaya
• Improve NVTX annotation for functions (#671) @kingcrimsontianyu
• Use latest release of cmake-maven-plugin (#652) @bdice

v25.04.00

🚨 Breaking Changes

• Improve config handling in cpp and python (#644) @kingcrimsontianyu
• Adjust the way of determining FileHandle's compatibility mode for sync and async I/O to improve code readability (#608) @kingcrimsontianyu

🐛 Bug Fixes

• Fix compilation error under openSUSE Tumbleweed (#667) @mrjoe7
• Another fix to update-version (#663) @raydouglass
• Add pom.xml to update-version.sh (#662) @raydouglass
• Fix an RPATH-related bug for CMake install (#661) @kingcrimsontianyu
• Minor fixes on the file permission, host allocation flag, and worker thread initialization (#637) @kingcrimsontianyu
• Fix package versioning (#616) @vyasr
• Fix std::future status query (#596) @kingcrimsontianyu

📖 Documentation

• add install docs (#638) @Intron7

🚀 New Features

• Add C++ benchmarks (part 1/n) (#659) @kingcrimsontianyu
• Enable tracing of thread pool tasks using NVTX (#630) @kingcrimsontianyu

🛠️ Improvements

• Add macros KVIKIO_EXPECT and KVIKIO_FAIL to improve exception handling (#653) @kingcrimsontianyu
• nvCOMP : don't use legacy API (#650) @madsbk
• Use conda-build instead of conda-mambabuild (#648) @bdice
• Remove FindCUDAToolkit.cmake from .pre-commit-config.yaml (#647) @KyleFromNVIDIA
• Support zarr-python 3.x (#646) @TomAugspurger
• Remove FindCUDAToolkit backport (#645) @KyleFromNVIDIA
• Improve config handling in cpp and python (#644) @kingcrimsontianyu
• Added timeout to kvikio requests (#643) @TomAugspurger
• Consolidate more Conda solves in CI (#636) @KyleFromNVIDIA
• Require CMake 3.30.4 (#635) @robertmaynard
• Use rapids-generate-version for package versions (#628) @bdice
• Forward-merge branch-25.02 into branch-25.04 (#626) @bdice
• Use shared-workflows branch-25.04 (#623) @bdice
• Add shellcheck to pre-commit and fix warnings (#621) @gforsyth
• Add build_type input field for test.yaml (#620) @gforsyth
• add telemetry (#615) @msarahan
• Add RAII file wrappers to avoid resource leak (#614) @kingcrimsontianyu
• disallow fallback to Make in Python builds (#612) @jameslamb
• Add verify-codeowners hook (#611) @KyleFromNVIDIA
• Forward-merge branch-25.02 to branch-25.04 (#610) @bdice
• Adjust the way of determining FileHandle's compatibility mode for sync and async I/O to improve code readability (#608) @kingcrimsontianyu
• Migrate to NVKS for amd64 CI runners (#605) @bdice
• Retry on HTTP 50x errors (#603) @TomAugspurger
• Add tutorials to showcase GDS for nifti and dicom images (#597) @yiheng-wang-nv
• Apply east const using clang-format. Fix header issues reported by clang-tidy (#594) @kingcrimsontianyu

v25.02.01

🚨 Breaking Changes

• Continue to make KvikIO a shared library by moving code from hpp to cpp (#581) @kingcrimsontianyu

🐛 Bug Fixes

• Set POSITION_INDEPENDENT_CODE on libcurl (#625) @KyleFromNVIDIA
• allow deselecting nvcomp wheels (#589) @jameslamb
• Fix the misplaced libcurl in CMake (#587) @kingcrimsontianyu

🚀 New Features

• Remove upper bounds on cuda-python to allow 12.6.2 and 11.8.5 (#564) @bdice

🛠️ Improvements

• Update to nvcomp 4.2.0.11 (#639) @bdice
• Use rapids-pip-retry in CI jobs that might need retries (#619) @gforsyth
• Use std::string (#617) @madsbk
• Revert CUDA 12.8 shared workflow branch changes (#609) @vyasr
• Build and test with CUDA 12.8.0 (#600) @bdice
• Separate more interface and definition. Add comments on std::future. Mark noexcept to compat mode-related functions (#588) @kingcrimsontianyu
• Use GCC 13 in CUDA 12 conda builds. (#586) @bdice
• Zarr: pin to version 2.* (#583) @madsbk
• Continue to make KvikIO a shared library by moving code from hpp to cpp (#581) @kingcrimsontianyu
• Check if nightlies have succeeded recently enough (#579) @vyasr
• Update cuda-python lower bounds to 12.6.2 / 11.8.5 (#573) @bdice
• Improve compatibility mode consistency (#567) @kingcrimsontianyu
• Require approval to run CI on draft PRs (#563) @bdice
• Add breaking change workflow trigger (#536) @AyodeAwe
• Add Initial Java Support for GDS to KvikIO (#396) @aslobodaNV

v24.12.00

🚨 Breaking Changes

• Use curl >=8.5.0 to align with conda-forge and avoid CVEs. (#574) @bdice
• cufile version (#565) @madsbk
• Add a new KvikIO compatibility mode "AUTO" (#547) @kingcrimsontianyu
• Build KvikIO as a shared library (#527) @madsbk
• Small improvements (#493) @kingcrimsontianyu

🐛 Bug Fixes

• Use curl >=8.5.0 to align with conda-forge and avoid CVEs. (#574) @bdice
• cufile version (#565) @madsbk
• Fix the pytest error for async io (#559) @kingcrimsontianyu
• fix library-loading issues in editable installs (#553) @jameslamb
• Backport FindCUDAToolkit from CMake 3.31 (#550) @KyleFromNVIDIA
• Fix exporting of include directories (#540) @KyleFromNVIDIA
• Moving details in file_handle.hpp to .cpp (#539) @madsbk
• Disallow cuda-python 12.6.1 and 11.8.4 (#537) @bdice
• Fix case of find_package call (#534) @vyasr
• CurlHandle: fix error message handling (#522) @madsbk
• Don't use macros for cuda driver functions (#516) @Jacobfaib
• Fix CUDA driver type stub definitions (#511) @Jacobfaib
• Stop tagging wheels as arch-agnostic (#507) @KyleFromNVIDIA
• Include <utility> since we use std::exchange (#505) @robertmaynard
• Disabling curl tests doesn't now disable kvikio cpp tests (#503) @robertmaynard
• Visibility of static class methods (#492) @madsbk

📖 Documentation

• to (#532) @RichardScottOZ

🚀 New Features

• Add a new KvikIO compatibility mode "AUTO" (#547) @kingcrimsontianyu
• Upgrade nvcomp to 4.1.0.6 (#525) @bdice

🛠️ Improvements

• prefer wheel-provided libkvikio.so, use RTLD_LOCAL (#551) @jameslamb
• enforce wheel size limits, README formatting in CI (#548) @jameslamb
• remove WheelHelpers.cmake (#545) @jameslamb
• Put a ceiling on cuda-python (#543) @jameslamb
• Replace FindcuFile with upstream FindCUDAToolkit support (#542) @KyleFromNVIDIA
• Build KvikIO as a shared library (#527) @madsbk
• print sccache stats in builds (#524) @jameslamb
• RemoteIO: use a pinned bounce buffer (#519) @madsbk
• Use registered strings for NVTX. Add more NVTX annotations. (#518) @kingcrimsontianyu
• Minor README fixes (#517) @Jacobfaib
• Python bindings to cuFileDriverOpen() and cuFileDriverClose() (#514) @madsbk
• Add CUDA libs in Python Conda, Consolidate Conda CI installs & use rapids-dask-dependency (#513) @jakirkham
• S3 benchmark: adding cudf-kvikio and cudf-fsspec (#509) @madsbk
• Use Cython's array to back Py_ssize_t[::1] (#504) @jakirkham
• Mark all of Array's nogil cdef functions as noexcept (#502) @jakirkham
• Simplify _to_string encoding of Python strs (#498) @jakirkham
• make conda installs in CI stricter (#495) @jameslamb
• Small improvements (#493) @kingcrimsontianyu
• Prune workflows based on changed files (#489) @KyleFromNVIDIA
• Remote IO: S3 support (#479) @madsbk
• Use nvcomp wheel instead of bundling nvcomp (#478) @KyleFromNVIDIA
• Touch ups to Array (#475) @jakirkham
• Remote IO: http support (#464) @madsbk

v24.10.00

🚨 Breaking Changes

• Remove now unnecessary variable (#466) @vyasr

🐛 Bug Fixes

• Fix no-cuda build (#484) @madsbk
• Add license to kvikio wheel (#481) @raydouglass
• Fix release-after-main bug in get_primary_cuda_context (#472) @madsbk
• Intentionally leak static CUDA resources to avoid crash (part 2) (#462) @kingcrimsontianyu
• Fix several issues with wheel builds (#439) @KyleFromNVIDIA
• Fix missing 24.08 => 24.10 renames (#423) @madsbk

📖 Documentation

• Docs cleanup (#480) @madsbk

🚀 New Features

• Update nvcomp to 4.0.1 (#449) @vuule

🛠️ Improvements

• nvtx support macos (#474) @madsbk
• Update update-version.sh to use packaging lib (#469) @AyodeAwe
• Remove now unnecessary variable (#466) @vyasr
• Use CI workflow branch 'branch-24.10' again (#465) @jameslamb
• Update to flake8 7.1.1. (#463) @bdice
• remove pytest dependency in kvikio conda recipe (#461) @jameslamb
• Add support for Python 3.12, relax numcodecs pin (#458) @jameslamb
• Skip cufile and nvtx3 when cuda isn't found (#455) @madsbk
• Non-static bounce buffer option (#454) @madsbk
• Use nvcomp defaults for algo options. (#450) @bdice
• Update rapidsai/pre-commit-hooks (#448) @KyleFromNVIDIA
• Make the bounce buffer size configurable (#447) @madsbk
• Update build.sh (#445) @madsbk
• Drop Python 3.9 support (#443) @jameslamb
• Remove NumPy <2 pin (#441) @seberg
• C++ Test (#440) @madsbk
• config_path(): return a reference (#438) @madsbk
• Stream sync before calling cufile (#437) @madsbk
• Speeding up Python tests (#435) @madsbk
• Cython refactoring (#434) @madsbk
• Update pre-commit hooks (#433) @KyleFromNVIDIA
• Improve update-version.sh (#432) @bdice
• Use tool.scikit-build.cmake.version, set scikit-build-core minimum-version (#430) @jameslamb
• build.sh: --pydevelop (#429) @madsbk