v24.10.00

🐛 Bug Fixes

• Ensure pylibwholegraph conda packages have the license (#215) @raydouglass
• fix_mnnvl_with_uuid (#207) @chuangz0

🛠️ Improvements

• bump NCCL floor to 2.19 (#223) @jameslamb
• Update update-version.sh to use packaging lib (#219) @AyodeAwe
• bump NCCL floor to 2.18.1.1, relax PyTorch pin (#218) @jameslamb
• Use CI workflow branch 'branch-24.10' again (#216) @jameslamb
• Add support for Python 3.12 (#214) @jameslamb
• Update rapidsai/pre-commit-hooks (#213) @KyleFromNVIDIA
• Drop Python 3.9 support (#209) @jameslamb
• Remove NumPy <2 pin (#208) @seberg
• Update pre-commit hooks (#206) @KyleFromNVIDIA
• Improve update-version.sh (#204) @bdice
• Use tool.scikit-build.cmake.version, set scikit-build-core minimum-version (#203) @jameslamb
• Add horovodrun launch agent for Wholegraph (#200) @Tomcli
• support different entry size for different ranks (#194) @zhuofan1123
• Remove Dockerfile (#184) @bdice

[NIGHTLY] v24.12.00

🔗 Links

• Development Branch
• Compare with main branch

🚨 Breaking Changes

• stop publishing packages (#233) @jameslamb

🛠️ Improvements

• stop publishing packages (#233) @jameslamb
• print sccache stats in builds (#231) @jameslamb
• manage more dependencies in dependencies.yaml, declare 'numpy' runtime dependency (#230) @jameslamb
• make conda installations in CI stricter (#228) @jameslamb
• Add a new memory type: Hierarchy (#227) @zhuofan1123
• Prune workflows based on changed files (#226) @KyleFromNVIDIA

v24.08.00

🐛 Bug Fixes

• Add cuda-nvml-dev to dependencies.yaml. (#197) @bdice
• Revert "Add CUDA_STATIC_MATH_LIBRARIES" (#192) @KyleFromNVIDIA
• fixed bugs (#180) @zhuofan1123

🛠️ Improvements

• clarify which dependencies in dependencies.yaml are conda-only (#195) @jameslamb
• Use workflow branch 24.08 again (#193) @KyleFromNVIDIA
• Build and test with CUDA 12.5.1 (#191) @KyleFromNVIDIA
• Add CUDA_STATIC_MATH_LIBRARIES (#190) @KyleFromNVIDIA
• skip CMake 3.30.0 (#189) @jameslamb
• Use verify-alpha-spec hook (#188) @KyleFromNVIDIA
• allow users to choose shm allocation method for chunked/continous host memory (#187) @linhu-nv
• decouple embedding creation from optimizer (#186) @zhuofan1123
• Mnnvl with split comm (#185) @chuangz0
• Adopt CI/packaging codeowners (#183) @bdice
• use rapids-build-backend (#181) @jameslamb

v24.06.00

🐛 Bug Fixes

• a quick fix to wholememory tensor gather default data type (#173) @linhu-nv
• quick fix to a map_indice bug && add comment for parameter round_robin_size (#172) @linhu-nv

🚀 New Features

• Add initial support of distributed sampling (#171) @chang-l

🛠️ Improvements

• Sort indices before gathering (#174) @zhuofan1123
• Always use a static gtest (#167) @vyasr
• Fix host view for mnnvl (#166) @chuangz0
• subwarp version gather op for small embedding size (#165) @chuangz0
• Migrate to {{ stdlib("c") }} (#164) @hcho3
• support read file with multi threads and add test_wholememory_io for round-roubin read (#163) @chuangz0
• fix CI issue due to pytorch and mkl version conflict (#162) @linhu-nv
• allow temp_memory_handler to allocate memory for multiple times (#161) @linhu-nv
• remove unnecessary sync between thrust ops and host threads (#160) @linhu-nv
• Remove scripts/checks/copyright.py (#149) @KyleFromNVIDIA

v24.04.00

🐛 Bug Fixes

• Update pre-commit-hooks to v0.0.3 (#152) @KyleFromNVIDIA
• Fixed README links to point to cuGraph API (#145) @acostadon
• [Bugfix] Fix to compile when NVSHMEM is ON (#142) @chang-l
• handle more RAPIDS version formats in update-version.sh (#122) @jameslamb

🚀 New Features

• Support CUDA 12.2 (#116) @jameslamb

🛠️ Improvements

• Use conda env create --yes instead of --force (#155) @bdice
• add round-robin shard strategy (#154) @linhu-nv
• Switch to scikit-build-core (#150) @vyasr
• Update script input name (#147) @AyodeAwe
• Add upper bound to prevent usage of NumPy 2 (#146) @bdice
• Replace local copyright check with pre-commit-hooks verify-copyright (#144) @KyleFromNVIDIA
• remove an unnecessary sync in exchange_embeddings_nccl_func (#143) @linhu-nv
• Use default rapids-cmake CUDA_ARCHITECTURES (#140) @trxcllnt
• Add support for Python 3.11, require NumPy 1.23+ (#139) @jameslamb
• [Bugfix] Host full-neighbor sampling returns wrong results in unit test (#138) @chang-l
• use enum to implement log_level in wholememory (#136) @linhu-nv
• target branch-24.04 for GitHub Actions workflows (#135) @jameslamb
• Add environment-agnostic scripts for running ctests and pytests (#128) @trxcllnt

v24.02.00

🐛 Bug Fixes

• Revert "Exclude tests from builds (#127)" (#130) @raydouglass
• Exclude tests from builds (#127) @vyasr
• fix a bug for embedding optimizer, which leads to undefined behavior (#108) @linhu-nv
• fix inferencesample option (#107) @chuangz0

🚀 New Features

• allow users to control gather/scatter sms (#124) @linhu-nv

🛠️ Improvements

• Logging level (#123) @linhu-nv
• Fix pip dependencies (#118) @trxcllnt
• Remove usages of rapids-env-update (#117) @KyleFromNVIDIA
• refactor CUDA versions in dependencies.yaml (#115) @jameslamb
• Don't overwrite wholegraph_ROOT if provided (#114) @vyasr
• added Direct IO support for WholeMemory loading (#113) @dongxuy04
• Align versions for cudnn, clang-tools, cython, and doxygen with the rest of RAPIDS. (#112) @bdice
• Reset WholeGraph communicators during the finalize call (#111) @chang-l
• Forward-merge branch-23.12 to branch-24.02 (#102) @bdice

v23.12.00

🐛 Bug Fixes

• move_vector_clear_outside_loop (#103) @chuangz0
• change pytorch cu121 to stable to fix ci (#97) @dongxuy04

🚀 New Features

• Integrate NVSHMEM into WholeGraph (#91) @chuangz0
• Grace Hopper support and add benchmark (#87) @dongxuy04

🛠️ Improvements

• Fix dependencies on librmm and libraft. (#96) @bdice
• gather/scatter optimizations (#90) @linhu-nv
• Use branch-23.12 workflows. (#84) @AyodeAwe
• Setup Consistent Nightly Versions for Pip and Conda (#82) @divyegala
• Add separate init, expose gather/scatter for WholeMemoryTensor and update example (#81) @dongxuy04
• Use RNG (random number generator) provided by RAFT (#79) @linhu-nv
• Build CUDA 12.0 ARM conda packages. (#74) @bdice
• upload xml docs (#73) @AyodeAwe
• replace optparse with argparser (#61) @chuangz0

v23.10.00

🐛 Bug Fixes

• Update all versions to 23.10 (#71) @raydouglass
• Use conda mambabuild not mamba mambabuild (#67) @bdice

🛠️ Improvements

• Update image names (#70) @AyodeAwe
• Update to clang 16.0.6. (#68) @bdice
• Simplify wheel build scripts and allow alphas of RAPIDS dependencies (#66) @divyegala
• Fix docs build and slightly optimize (#63) @dongxuy04
• Use copy-pr-bot (#60) @ajschmidt8
• PR: Use top-k from RAFT (#53) @chuangz0

v23.08.01

🚨 Breaking Changes

• Refactoring into 23.08 (#24) @BradReesWork

🐛 Bug Fixes

• Add LICENSE to wheels (#55) @raydouglass

🛠️ Improvements

• Correct syntax in GHA workflow (#46) @tingyu66
• Refactoring into 23.08 (#24) @BradReesWork

v23.08.00

🚨 Breaking Changes

• Refactoring into 23.08 (#24) @BradReesWork

🛠️ Improvements

• Correct syntax in GHA workflow (#46) @tingyu66
• Refactoring into 23.08 (#24) @BradReesWork