v26.10.00

What's Changed

🚨 Breaking Changes

• FastIntDiv: fallback to normal division when values exceed 32-bit ran… by @huuanhhuyn in #3093
• Better error reporting with std::source_location by @achirkin in #3121
• Migrate stream APIs from rmm::cuda_stream_view to cuda::stream_ref by @bdice in #3129

🐛 Bug Fixes

• Fix missing nvtx stack and host mem resource by exporting the symbols by @huuanhhuyn in #3083
• Update pre-commit-hooks, fix UCXX version scheme by @KyleFromNVIDIA in #3087
• Remove orphaned raft::runtime::matrix::select_k declaration by @says1117 in #3084
• Make cuBLASLt descriptor wrappers move-safe by @fallintoplace in #3078
• Mitigate cuBLASLt 13.6 GEMM bug for inputs greater than 2^31 by @divyegala in #3098
• cublas team verified workaround for large GEMM algo68 bug by @divyegala in #3100
• Restore tracked device resources on their original GPU by @fallintoplace in #3077
• Fix lanczos launch grid by @achirkin in #3109
• Fix launch_kernel handling of restricted and partially specified kernel arguments by @dantegd in #3117
• Fix devcontainer cache version updates by @bdice in #3119
• Use l-values to initialize RMM constructors by @divyegala in #3123
• Fix curand stream ordering in MST alteration by @dantegd in #3125
• Force stream sync in raft::device_reference by @achirkin in #3124
• Synchronize MST edge count update by @bdice in #3130
• Fix strided reduction block count by @achirkin in #3127
• Fix striding in mdspan 1d copies by @divyegala in #3134
• Update UCX calls to use new send/recv functions by @ChuckHastings in #3135

📖 Documentation

• Enable public docs features in CI by @bdice in #3131

🚀 New Features

• Add sparse Lanczos SVD solver by @Intron7 in #3034
• [REVIEW] Addition of gemm_strided_batched by @landrumb in #3106
• Implement a kernel dispatcher raft::launch_kernel by @achirkin in #3104
• Dry Run Protocol by @achirkin in #2961

🛠️ Improvements

• Update RAPIDS.cmake to log source of rapids-cmake by @arhag23 in #3074
• Empty commit to trigger a build by @bdice in #3089
• Use cuda::std::bit_cast in stats minmax instead of hand-rolled helper by @says1117 in #3095
• Use cuda::std::numeric_limits in LAP kernels instead of passing infinity by @says1117 in #3094
• enforce 'yamllint' checks by @jameslamb in #3096
• ensure nightly builds always produce new packages, expand 'changed-files' lists by @jameslamb in #3103
• Update to rapids-logger 0.3 by @bdice in #3102
• wheels: enforce 'abi3audit' checks by @jameslamb in #3111
• ci: avoid triggered-by-schedule and triggered-by-merge builds cancelling each other by @jameslamb in #3120
• X-ORG-3107: Publish API docs to docs.nvidia.com by @josephine-wolf-oberholtzer in #3108
• Audit Wheel Dependencies by @tarang-jain in #3091
• Extend kernel_launch with kernel_ref and attributes by @achirkin in #3128
• [FEA] permute: replace LCG with keyed Feistel permutation, add deterministic key API by @vinaydes in #3079
• Replace hardcoded Lanczos eigenvalue goldens with property-based checks by @says1117 in #3086
• Backport: Relax lanczos gtest tolerance by @bdice in #3142
• Adopt CUDA stream compatibility accessors by @bdice in #3136
• Discover CMake packages installed by wheels by @bdice in #3143
• Add stream pool configuration to device_resources_snmg by @viclafargue in #3149
• Relax lanczos gtest threshold by @aamijar in #3153
• X-ORG-410: Integrate archived docs into version switcher by @josephine-wolf-oberholtzer in #3162

New Contributors

• @arhag23 made their first contribution in #3074
• @says1117 made their first contribution in #3084
• @fallintoplace made their first contribution in #3078
• @landrumb made their first contribution in #3106

Full Changelog: v26.10.00a...v26.10.00