v25.06.00

🚨 Breaking Changes

• Decoupling multi gpu resources from nccl usage (#2647) @jinsolp

🐛 Bug Fixes

• NCCL comm resource fix (#2692) @viclafargue
• Fix the launch bounds for nn-descent kernel for 1210 and remove nn-descent tests (#2691) @viclafargue
• Prefer host gather when dataset is available both on host and device (#2671) @tfeher
• Fix warnings treated as errors downstream in cuVS (#2644) @achirkin
• Fix nccl_comm.hpp warning: #83-D: type qualifier specified more than once (#2643) @achirkin
• NVTX: null destination pointer warning-treated-as-error (#2639) @achirkin
• Add UCXX and NCCL to libraft conda recipe (#2636) @divyegala
• Fix building cutlass (#2619) @miscco
• Fix COO symmetrization (#2582) @viclafargue

🚀 New Features

• [Feat] add cudaMemcpy2DAsync wrapper (#2674) @rhdong
• Python wrapper for device_resources_snmg (#2666) @jinsolp
• Laplacian normalization primitives (#2648) @aamijar
• [FEA] Matrix shift rows and columns (#2634) @jinsolp
• Use NCCL wheels from PyPI for CUDA 12 builds (#2629) @divyegala
• Support strided matrix view as an input to matrix::samples_rows (#2626) @enp1s0
• [Feat] add support for bm25 and tfidf (#2567) @jperez999

🛠️ Improvements

• use 'rapids-init-pip' in wheel CI, other CI changes (#2677) @jameslamb
• Dask 2025.4.1 compatibility (#2673) @TomAugspurger
• Finish CUDA 12.9 migration and use branch-25.06 workflows (#2669) @bdice
• Update to clang 20 (#2665) @bdice
• Quote head_rev in conda recipes (#2660) @bdice
• CUDA 12.9 use updated compression flags (#2657) @robertmaynard
• Build and test with CUDA 12.9.0 (#2655) @bdice
• Exclude librmm.so from auditwheel (#2654) @bdice
• Fix cub include in normalize.cuh (#2652) @lowener
• Add support for Python 3.13 (#2649) @gforsyth
• Decoupling multi gpu resources from nccl usage (#2647) @jinsolp
• [BUGFIX] Fixed quoting in wheel paths in pylibraft and raft_dask wheel tests (#2645) @VenkateshJaya
• Download build artifacts from Github for CI (#2640) @VenkateshJaya
• Limit allowed wheel sizes (#2638) @divyegala
• Remove CUDA whole compilation ODR violations (#2633) @divyegala
• refactor(rattler): enable strict channel priority for builds (#2632) @gforsyth
• Vendor RAPIDS.cmake (#2631) @bdice
• Replace Thrust iterator facilities and replace them with libcu++ ones (#2627) @miscco
• Port all conda recipes to rattler-build (#2623) @gforsyth
• Add missing thrust include (#2618) @miscco
• Moving wheel builds to specified location and uploading build artifacts to Github (#2617) @VenkateshJaya
• Fixed pytest marker warnings by removing unused pytest.ini (#2591) @TomAugspurger
• Introduction of the raft::device_resources_snmg type (#2549) @viclafargue
• Create a NCCL sub-communicator using ncclCommSplit (#2495) @seunghwak