rmm

🚨 Breaking Changes

• Accept stream argument in DeviceMemoryResource allocate/deallocate (#1494) @wence-
• Replace all internal usage of get_upstream with get_upstream_resource (#1491) @miscco
• Deprecate rmm::mr::device_memory_resource::supports_streams() (#1452) @harrism
• Remove deprecated rmm::detail::available_device_memory (#1438) @harrism
• Make device_memory_resource::supports_streams() not pure virtual. Remove derived implementations and calls in RMM (#1437) @harrism
• Deprecate rmm::mr::device_memory_resource::get_mem_info() and supports_get_mem_info(). (#1436) @harrism

🐛 Bug Fixes

• Fix search path for torch allocator in editable installs and ensure CUDA support is available (#1498) @vyasr
• Accept stream argument in DeviceMemoryResource allocate/deallocate (#1494) @wence-
• Run STATISTICS_TEST and TRACKING_TEST in serial to avoid OOM errors. (#1487) @bdice

📖 Documentation

• Pin to recent breathe, to prevent getting an unsupported sphinx version. (#1495) @bdice

🚀 New Features

• Replace all internal usage of get_upstream with get_upstream_resource (#1491) @miscco
• Add complete set of resource ref aliases (#1479) @nvdbaranec
• Automate include grouping using clang-format (#1463) @harrism
• Add get_upstream_resource to resource adaptors (#1456) @miscco
• Deprecate rmm::mr::device_memory_resource::supports_streams() (#1452) @harrism
• Remove duplicated memory_resource_tests (#1451) @miscco
• Change rmm::exec_policy to take async_resource_ref (#1449) @miscco
• Change device_scalar to take async_resource_ref (#1447) @miscco
• Add device_async_resource_ref convenience alias (#1441) @harrism
• Remove deprecated rmm::detail::available_device_memory (#1438) @harrism
• Make device_memory_resource::supports_streams() not pure virtual. Remove derived implementations and calls in RMM (#1437) @harrism
• Deprecate rmm::mr::device_memory_resource::get_mem_info() and supports_get_mem_info(). (#1436) @harrism
• Support CUDA 12.2 (#1419) @jameslamb

🛠️ Improvements

• Use conda env create --yes instead of --force (#1509) @bdice
• Add upper bound to prevent usage of NumPy 2 (#1501) @bdice
• Remove hard-coding of RAPIDS version where possible (#1496) @KyleFromNVIDIA
• Requre NumPy 1.23+ (#1488) @jakirkham
• Use rmm::device_async_resource_ref in multi_stream_allocation benchmark (#1482) @miscco
• Update devcontainers to CUDA Toolkit 12.2 (#1470) @trxcllnt
• Add support for Python 3.11 (#1469) @jameslamb
• target branch-24.04 for GitHub Actions workflows (#1468) @jameslamb
• [FEA]: Use std::optional instead of thrust::optional (#1464) @miscco
• Add environment-agnostic scripts for running ctests and pytests (#1462) @trxcllnt
• Ensure that ctest is called with --no-tests=error. (#1460) @bdice
• Update ops-bot.yaml (#1458) @AyodeAwe
• Adopt the rmm::device_async_resource_ref alias (#1454) @miscco
• Refactor error.hpp out of detail (#1439) @lamarrr

🚨 Breaking Changes

• Make device_memory_resource::do_get_mem_info() and supports_get_mem_info() not pure virtual. Remove derived implementations and calls in RMM (#1430) @harrism
• Deprecate detail::available_device_memory, most detail/aligned.hpp utilities, and optional pool_memory_resource initial size (#1424) @harrism
• Require explicit pool size in pool_memory_resource and move some things out of detail namespace (#1417) @harrism
• Remove HTML builds of librmm (#1415) @vyasr
• Update to CCCL 2.2.0. (#1404) @bdice
• Switch to scikit-build-core (#1287) @vyasr

🐛 Bug Fixes

• Exclude tests from builds (#1459) @vyasr
• Update CODEOWNERS (#1410) @raydouglass
• Correct signatures for torch allocator plug in (#1407) @wence-
• Fix Arena MR to support simultaneous access by PTDS and other streams (#1395) @tgravescs
• Fix else-after-throw clang tidy error (#1391) @harrism

📖 Documentation

• remove references to setup.py in docs (#1420) @jameslamb
• Remove HTML builds of librmm (#1415) @vyasr
• Update GPU support docs to drop Pascal (#1413) @harrism

🚀 New Features

• Make device_memory_resource::do_get_mem_info() and supports_get_mem_info() not pure virtual. Remove derived implementations and calls in RMM (#1430) @harrism
• Deprecate detail::available_device_memory, most detail/aligned.hpp utilities, and optional pool_memory_resource initial size (#1424) @harrism
• Add a host-pinned memory resource that can be used as upstream for pool_memory_resource. (#1392) @harrism

🛠️ Improvements

• Remove usages of rapids-env-update (#1423) @KyleFromNVIDIA
• Refactor CUDA versions in dependencies.yaml. (#1422) @bdice
• Require explicit pool size in pool_memory_resource and move some things out of detail namespace (#1417) @harrism
• Update dependencies.yaml to support CUDA 12.*. (#1414) @bdice
• Define python dependency range as a matrix fallback. (#1409) @bdice
• Use latest cuda-python within CUDA major version. (#1406) @bdice
• Update to CCCL 2.2.0. (#1404) @bdice
• Remove RMM_BUILD_WHEELS and standardize Python builds (#1401) @vyasr
• Update to fmt 10.1.1 and spdlog 1.12.0. (#1374) @bdice
• Switch to scikit-build-core (#1287) @vyasr

🚨 Breaking Changes

• Document minimum CUDA version of 11.4 (#1385) @harrism
• Store and set the correct CUDA device in device_buffer (#1370) @harrism
• Use cuda::mr::memory_resource instead of raw device_memory_resource (#1095) @miscco

🐛 Bug Fixes

• Update actions/labeler to v4 (#1397) @raydouglass
• Backport arena MR fix for simultaneous access by PTDS and other streams (#1396) @bdice
• Deliberately leak PTDS thread_local events in stream ordered mr (#1375) @wence-
• Add missing CUDA 12 dependencies and fix dlopen library names (#1366) @vyasr

📖 Documentation

• Document minimum CUDA version of 11.4 (#1385) @harrism
• Fix more doxygen issues (#1367) @vyasr
• Add groups to the doxygen docs (#1358) @vyasr
• Enable doxygen XML and fix issues (#1348) @vyasr

🚀 New Features

• Make internally stored default argument values public (#1373) @vyasr
• Store and set the correct CUDA device in device_buffer (#1370) @harrism
• Update rapids-cmake functions to non-deprecated signatures (#1357) @robertmaynard
• Generate unified Python/C++ docs (#1324) @vyasr
• Use cuda::mr::memory_resource instead of raw device_memory_resource (#1095) @miscco

🛠️ Improvements

• Silence false gcc warning (#1381) @miscco
• Build concurrency for nightly and merge triggers (#1380) @bdice
• Update shared-action-workflows references (#1363) @AyodeAwe
• Use branch-23.12 workflows. (#1360) @bdice
• Update devcontainers to 23.12 (#1355) @raydouglass
• Generate proper, consistent nightly versions for pip and conda packages (#1347) @vyasr
• RMM: Build CUDA 12.0 ARM conda packages. (#1330) @bdice

🚨 Breaking Changes

• Update to Cython 3.0.0 (#1313) @vyasr

🐛 Bug Fixes

• Compile cdef public functions from torch_allocator with C ABI (#1350) @wence-
• Make doxygen only a conda dependency. (#1344) @bdice
• Use conda mambabuild not mamba mambabuild (#1338) @wence-
• Fix stream_ordered_memory_resource attempt to record event in stream from another device (#1333) @harrism

📖 Documentation

• Clean up headers in CMakeLists.txt. (#1341) @bdice
• Add pre-commit hook to validate doxygen (#1334) @vyasr
• Fix doxygen warnings (#1317) @vyasr
• Treat warnings as errors in Python documentation (#1316) @vyasr

🚀 New Features

• Enable RMM Debug Logging via Python (#1339) @harrism

🛠️ Improvements

• Update image names (#1346) @AyodeAwe
• Update to clang 16.0.6. (#1343) @bdice
• Update doxygen to 1.9.1 (#1337) @vyasr
• Simplify wheel build scripts and allow alphas of RAPIDS dependencies (#1335) @divyegala
• Use copy-pr-bot (#1329) @ajschmidt8
• Add RMM devcontainers (#1328) @trxcllnt
• Add Python bindings for limiting_resource_adaptor (#1327) @pentschev
• Fix missing jQuery error in docs (#1321) @AyodeAwe
• Use fetch_rapids.cmake. (#1319) @bdice
• Update to Cython 3.0.0 (#1313) @vyasr
• Branch 23.10 merge 23.08 (#1312) @vyasr
• Branch 23.10 merge 23.08 (#1309) @vyasr

🚨 Breaking Changes

• Stop invoking setup.py (#1300) @vyasr
• Remove now-deprecated top-level allocator functions (#1281) @wence-
• Remove padding from device_memory_resource (#1278) @vyasr

🐛 Bug Fixes

• Fix typo in wheels-test.yaml. (#1310) @bdice
• Add a missing '#include <array>' in logger.hpp (#1295) @valgur
• Use gbench thread_index() accessor to fix replay bench compilation (#1293) @harrism
• Ensure logger tests don't generate temp directories in build dir (#1289) @robertmaynard

🚀 New Features

• Remove now-deprecated top-level allocator functions (#1281) @wence-

🛠️ Improvements

• Switch to new CI wheel building pipeline (#1305) @vyasr
• Revert CUDA 12.0 CI workflows to branch-23.08. (#1303) @bdice
• Update linters: remove flake8, add ruff, update cython-lint (#1302) @vyasr
• Adding identify minimum version requirement (#1301) @hyperbolic2346
• Stop invoking setup.py (#1300) @vyasr
• Use cuda-version to constrain cudatoolkit. (#1296) @bdice
• Update to CMake 3.26.4 (#1291) @vyasr
• use rapids-upload-docs script (#1288) @AyodeAwe
• Reorder parameters in RMM_EXPECTS (#1286) @vyasr
• Remove documentation build scripts for Jenkins (#1285) @ajschmidt8
• Remove padding from device_memory_resource (#1278) @vyasr
• Unpin scikit-build upper bound (#1275) @vyasr
• RMM: Build CUDA 12 packages (#1223) @bdice

🚨 Breaking Changes

• Update minimum Python version to Python 3.9 (#1252) @shwina

🐛 Bug Fixes

• Ensure Logger tests aren't run in parallel (#1277) @robertmaynard
• Pin to scikit-build<0.17.2. (#1262) @bdice

🛠️ Improvements

• Require Numba 0.57.0+ & NumPy 1.21.0+ (#1279) @jakirkham
• Align test_cpp.sh with conventions in other RAPIDS repos. (#1269) @bdice
• Switch back to using primary shared-action-workflows branch (#1268) @vyasr
• Update recipes to GTest version >=1.13.0 (#1263) @bdice
• Support CUDA 12.0 for pip wheels (#1259) @bdice
• Add build vars (#1258) @AyodeAwe
• Enable sccache hits from local builds (#1257) @AyodeAwe
• Revert to branch-23.06 for shared-action-workflows (#1256) @shwina
• run docs builds nightly too (#1255) @AyodeAwe
• Build wheels using new single image workflow (#1254) @vyasr
• Update minimum Python version to Python 3.9 (#1252) @shwina
• Remove usage of rapids-get-rapids-version-from-git (#1251) @jjacobelli
• Remove wheel pytest verbosity (#1249) @sevagh
• Update clang-format to 16.0.1. (#1246) @bdice
• Remove uses-setup-env-vars (#1242) @vyasr
• Move RMM_LOGGING_ASSERT into separate header (#1241) @ahendriksen
• Use ARC V2 self-hosted runners for GPU jobs (#1239) @jjacobelli

🐛 Bug Fixes

• Remove MANIFEST.in use auto-generated one for sdists and package_data for wheels (#1233) @vyasr
• Fix update-version.sh. (#1227) @vyasr
• Specify include_package_data to setup (#1218) @vyasr
• Revert changes overriding rapids-cmake repo. (#1209) @bdice
• Synchronize stream in DeviceBuffer.c_from_unique_ptr constructor (#1100) @shwina

🚀 New Features

• Use rapids-cmake parallel testing feature (#1183) @robertmaynard

🛠️ Improvements

• Stop setting package version attribute in wheels (#1236) @vyasr
• Add codespell as a linter (#1231) @bdice
• Pass AWS_SESSION_TOKEN and SCCACHE_S3_USE_SSL vars to conda build (#1230) @ajschmidt8
• Update to GCC 11 (#1228) @bdice
• Fix some minor oversights in the conversion to pyproject.toml (#1226) @vyasr
• Remove pickle compatibility layer in tests for Python < 3.8. (#1224) @bdice
• Move external allocators into rmm.allocators module to defer imports (#1221) @wence-
• Generate pyproject.toml dependencies using dfg (#1219) @vyasr
• Run rapids-dependency-file-generator via pre-commit (#1217) @vyasr
• Skip docs job in nightly runs (#1215) @AyodeAwe
• CI: Remove specification of manual stage for check_style.sh script. (#1214) @csadorf
• Use script rather than environment variable to modify package names (#1212) @vyasr
• Reduce error handling verbosity in CI tests scripts (#1204) @AjayThorve
• Update shared workflow branches (#1203) @ajschmidt8
• Use date in build string instead of in the version. (#1195) @bdice
• Stop using versioneer to manage versions (#1190) @vyasr
• Update to spdlog>=1.11.0, fmt>=9.1.0. (#1177) @bdice
• Migrate as much as possible to pyproject.toml (#1151) @jakirkham

🐛 Bug Fixes

• pre-commit: Update isort version to 5.12.0 (#1197) @wence-
• Revert "Upgrade to spdlog 1.10 (#1173)" (#1176) @bdice
• Ensure UpstreamResourceAdaptor is not cleared by the Python GC (#1170) @shwina

📖 Documentation

• Fix documentation author (#1188) @bdice

🚀 New Features

• Add RMM PyTorch allocator (#1168) @shwina

🛠️ Improvements

• Update shared workflow branches (#1201) @ajschmidt8
• Fix update-version.sh (#1199) @raydouglass
• Use CTK 118/cp310 branch of wheel workflows (#1193) @sevagh
• Update build.yaml workflow to reduce verbosity (#1192) @AyodeAwe
• Fix build.yaml workflow (#1191) @ajschmidt8
• add docs_build step (#1189) @AyodeAwe
• Upkeep/wheel param cleanup (#1187) @sevagh
• Update workflows for nightly tests (#1186) @ajschmidt8
• Build CUDA 11.8 and Python 3.10 Packages (#1184) @ajschmidt8
• Build wheels alongside conda CI (#1182) @sevagh
• Update conda recipes. (#1180) @bdice
• Update PR Workflow (#1174) @ajschmidt8
• Upgrade to spdlog 1.10 (#1173) @kkraus14
• Enable codecov (#1171) @ajschmidt8
• Add support for Python 3.10. (#1166) @bdice
• Update pre-commit hooks (#1154) @bdice

🐛 Bug Fixes

• Ensure consistent spdlog dependency target no matter the source (#1101) @robertmaynard
• Remove cuda event deadlocking issues in device mr tests (#1097) @robertmaynard
• Propagate exceptions raised in Python callback functions (#1096) @madsbk
• Avoid unused parameter warnings in do_get_mem_info (#1084) @fkallen
• Use rapids-cmake 22.10 best practice for RAPIDS.cmake location (#1083) @robertmaynard

📖 Documentation

• Document that minimum required CMake version is now 3.23.1 (#1098) @robertmaynard
• Fix docs for module-level API (#1091) @bdice
• Improve DeviceBuffer docs. (#1090) @bdice
• Branch 22.10 merge 22.08 (#1089) @harrism
• Improve docs formatting and update links. (#1086) @bdice
• Add resources section to README. (#1085) @bdice
• Simplify PR template. (#1080) @bdice

🚀 New Features

• Add gdb pretty-printers for rmm types (#1088) @upsj
• Support using THRUST_WRAPPED_NAMESPACE (#1077) @robertmaynard

🛠️ Improvements

• GH Actions - Enforce checks before builds run (#1125) @ajschmidt8
• Update GH Action Workflows (#1123) @ajschmidt8
• Add cudatoolkit versions to dependencies.yaml (#1119) @ajschmidt8
• Remove rmm installation from librmm tests` (#1117) @ajschmidt8
• Add GitHub Actions workflows (#1104) @Ethyling
• build.sh: accept --help (#1093) @madsbk
• Move clang dependency to conda develop packages. (#1092) @bdice
• Add device_uvector::reserve and device_buffer::reserve (#1079) @upsj
• Bifurcate Dependency Lists (#1073) @ajschmidt8

🐛 Bug Fixes

• Specify language as 'en' instead of None (#1059) @jakirkham
• Add a missed except * (#1057) @shwina
• Properly handle cudaMemHandleTypeNone and cudaErrorInvalidValue in is_export_handle_type_supported (#1055) @gerashegalov

📖 Documentation

• Centralize common css & js code in docs (#1075) @galipremsagar

🛠️ Improvements

• Add the ability to register and unregister reinitialization hooks (#1072) @shwina
• Update isort to 5.10.1 (#1069) @vyasr
• Forward merge 22.06 into 22.08 (#1067) @vyasr
• Forward merge 22.06 into 22.08 (#1066) @vyasr
• Pin max version of cuda-python to 11.7 (#1062) @galipremsagar
• Change build.sh to find C++ library by default and avoid shadowing CMAKE_ARGS (#1053) @vyasr

v22.06.01

🐛 Bug Fixes

• Clarifies Python requirements and version constraints (#1037) @jakirkham
• Use lib (not lib64) for libraries (#1024) @jakirkham
• Properly enable Cython docstrings. (#1020) @vyasr
• Update RMMNumbaManager to handle NUMBA_CUDA_USE_NVIDIA_BINDING=1 (#1004) @brandon-b-miller

📖 Documentation

• Clarify using RMM with other Python libraries (#1034) @jrhemstad
• Replace to_device with DeviceBuffer.to_device (#1033) @wence-
• Documentation Fix: Replace cudf::logic_error with rmm::logic_error (#1021) @codereport

🚀 New Features

• Add rmm::exec_policy_nosync (#1009) @fkallen
• Callback memory resource (#980) @shwina

🛠️ Improvements

• Fix conda recipes for conda compilers (#1043) @Ethyling
• Use new rapids-cython component of rapids-cmake to simplify builds (#1031) @vyasr
• Merge branch-22.04 to branch-22.06 (#1028) @jakirkham
• Update CMake pinning to just avoid 3.23.0. (#1023) @vyasr
• Build python using conda in GPU jobs (#1017) @Ethyling
• Remove pip requirements file. (#1015) @bdice
• Clean up Thrust includes. (#1011) @bdice
• Update black version (#1010) @vyasr
• Update cmake-format version for pre-commit and environments. (#995) @vyasr
• Use conda compilers (#977) @Ethyling
• Build conda packages using mambabuild (#900) @Ethyling

v22.04.01

🐛 Bug Fixes

• Add cuda-python dependency to pyproject.toml (#994) @sevagh
• Disable opportunistic reuse in async mr when cuda driver < 11.5 (#993) @rongou
• Use CUDA 11.2+ features via dlopen (#990) @robertmaynard
• Skip async mr tests when cuda runtime/driver < 11.2 (#986) @rongou
• Fix warning/error in debug assertion in device_uvector.hpp (#979) @harrism
• Fix signed/unsigned comparison warning (#970) @jlowe
• Fix comparison of async MRs with different underlying pools. (#965) @harrism

🚀 New Features

• Use scikit-build for the build process (#976) @vyasr

🛠️ Improvements

• Temporarily disable new ops-bot functionality (#1005) @ajschmidt8
• Rename librmm_tests to librmm-tests (#1000) @ajschmidt8
• Update librmm conda recipe (#997) @ajschmidt8
• Remove no_cma/has_cma variants (#996) @ajschmidt8
• Fix free-before-alloc in multithreaded test (#992) @aladram
• Add .github/ops-bot.yaml config file (#991) @ajschmidt8
• Log allocation failures (#988) @rongou
• Update librmm conda outputs (#983) @ajschmidt8
• Bump Python requirements in setup.cfg and rmm_dev.yml (#982) @shwina
• New benchmark compares concurrent throughput of device_vector and device_uvector (#981) @harrism
• Update librmm recipe to output librmm_tests package (#978) @ajschmidt8
• Update upload.sh to use --croot (#975) @AyodeAwe
• Fix conda uploads (#974) @ajschmidt8
• Add CMake install rules for tests (#969) @ajschmidt8
• Add device_buffer::ssize() and device_uvector::ssize() (#966) @harrism
• Added yml file for cudatoolkit version 11.6 (#964) @alhad-deshpande
• Replace ccache with sccache (#963) @ajschmidt8
• Make pool_memory_resource::pool_size() public (#962) @shwina
• Allow construction of cuda_async_memory_resource from existing pool (#889) @fkallen

🐛 Bug Fixes

• Use numba to get CUDA runtime version. (#946) @bdice
• Temporarily disable warnings for unknown pragmas (#942) @harrism
• Build benchmarks in RMM CI (#941) @harrism
• Headers that use std::thread now include <thread> (#938) @robertmaynard
• Fix failing stream test with a debug-only death test (#934) @harrism
• Prevent DeviceBuffer DeviceMemoryResource premature release (#931) @viclafargue
• Fix failing tracking test (#929) @harrism

🛠️ Improvements

• Prepare upload scripts for Python 3.7 removal (#952) @Ethyling
• Fix imports tests syntax (#935) @Ethyling
• Remove IncludeCategories from .clang-format (#933) @codereport
• Replace use of custom CUDA bindings with CUDA-Python (#930) @shwina
• Remove setup.py from update-release.sh script (#926) @ajschmidt8
• Improve C++ Test Coverage (#920) @harrism
• Improve the Arena allocator to reduce memory fragmentation (#916) @rongou
• Simplify CMake linting with cmake-format (#913) @vyasr

🚨 Breaking Changes

• Parameterize exception type caught by failure_callback_resource_adaptor (#898) @harrism

🐛 Bug Fixes

• Update recipes for Enhanced Compatibility (#910) @ajschmidt8
• Fix librmm uploads (#909) @ajschmidt8
• Use spdlog/fmt/ostr.h as it supports external fmt library (#907) @robertmaynard
• Fix variable names in logging macro calls (#897) @harrism
• Keep rapids cmake version in sync (#876) @robertmaynard

📖 Documentation

• Replace to_device() in docs with DeviceBuffer.to_device() (#902) @shwina
• Fix return value docs for supports_get_mem_info (#884) @harrism

🚀 New Features

• Out-of-memory callback resource adaptor (#892) @madsbk

🛠️ Improvements

• suppress spurious clang-tidy warnings in debug macros (#914) @rongou
• C++ code coverage support (#905) @harrism
• Provide ./build.sh flag to control CUDA async malloc support (#901) @robertmaynard
• Parameterize exception type caught by failure_callback_resource_adaptor (#898) @harrism
• Throw rmm::out_of_memory when we know for sure (#894) @rongou
• Update conda recipes for Enhanced Compatibility effort (#893) @ajschmidt8
• Add functions to query the stream of device_uvector and device_scalar (#887) @fkallen
• Add spdlog to install export set (#886) @trxcllnt

v21.10.01

🚨 Breaking Changes

• Delete cuda_async_memory_resource copy/move ctors/operators (#860) @jrhemstad

🐛 Bug Fixes

• Fix parameter name in asserts (#875) @vyasr
• Disallow zero-size stream pools (#873) @harrism
• Correct namespace usage in host memory resources (#872) @divyegala
• fix race condition in limiting resource adapter (#869) @rongou
• Install the right cudatoolkit in the conda env in gpu/build.sh (#864) @shwina
• Disable copy/move ctors and operator= from free_list classes (#862) @harrism
• Delete cuda_async_memory_resource copy/move ctors/operators (#860) @jrhemstad
• Improve concurrency of stream_ordered_memory_resource by stealing less (#851) @harrism
• Use the new RAPIDS.cmake to fetch rapids-cmake (#838) @robertmaynard

📖 Documentation

• Forward-merge branch-21.08 to branch-21.10 (#846) @jakirkham

🛠️ Improvements

• Forward-merge branch-21.08 into branch-21.10 (#877) @ajschmidt8
• Add .clang-tidy and fix clang-tidy warnings (#857) @harrism
• Update to use rapids-cmake 21.10 pre-configured packages (#854) @robertmaynard
• Clean up: use std::size_t, include cstddef and aligned.hpp where missing (#852) @harrism
• tweak the arena mr to reduce fragmentation (#845) @rongou
• Fix transitive include in cuda_device header (#843) @wphicks
• Refactor cmake style (#842) @robertmaynard
• add multi stream allocations benchmark. (#841) @cwharris
• Enforce default visibility for get_map. (#833) @trivialfis
• ENH Replace gpuci_conda_retry with gpuci_mamba_retry (#823) @dillon-cullinan
• Execution policy class (#816) @viclafargue

v21.08.02

v21.08.01