v26.08.00

What's Changed

🚨 Breaking Changes

• Switch RandomForest max_depth default to None by @csadorf in #8200
• Link to libcuvs dynamically in wheel builds by @divyegala in #8250
• Cleanup pairwise distances by @jcrist in #8268
• Deprecate convert_dtype by @jcrist in #8300
• Deprecate PorterStemmer by @jcrist in #8304
• Move Lars out of experimental by @jcrist in #8336
• CI Extend time limit for hypothesis tests by @betatim in #8346
• Deprecate cuml.tsa by @csadorf in #8344
• wheels: build CUDA 13 wheels with latest CTK (13.3.0) by @jameslamb in #8345
• Deprecate cuml.datasets.make_arima by @jcrist in #8379
• Deprecate legacy output types by @jcrist in #8389

🐛 Bug Fixes

• Fix pandas 3 string dtype compatibility and CI flakes by @csadorf in #8142
• Fix KDE score_samples symbol exports by @csadorf in #8173
• Fix hinge loss decision-function shape handling by @csadorf in #8245
• Fix clang-tidy failures in CUDA 13.2 CI by @dantegd in #8251
• Match sklearn random forest feature sampling by @RAMitchell in #8239
• Test cuml.accel upstream test suite with sklearn 1.9 by @csadorf in #8207
• Update test suite to work with new pytest release by @dantegd in #8266
• Introduce Safe Arithmetic by @divyegala in #8171
• Cleanup extract_knn_graph by @jcrist in #8272
• Fix RandomForestClassifier sklearn child metadata by @csadorf in #8291
• Fallback unsupported cuml.accel PCA and KernelRidge cases to CPU by @csadorf in #8290
• Fix FastIntDiv tparam by @aamijar in #8299
• Remove sklearn example xfails by @csadorf in #8090
• FIX: handle None transformers in ColumnTransformer get_params/set_params by @Husnain987 in #8292
• Handle OpenML checksum failures in sklearn examples by @csadorf in #8359
• Fix RF feature importances with infinite metrics by @RAMitchell in #8347
• Limit cuml.accel pytest logging to errors by @csadorf in #8360
• Fix --print-status path in benchmark CLI by @fallintoplace in #8320
• fix(clang-format): preserve SPDX comments by @bdice in #8368
• Make LabelEncoder docstring account for cuDF update by @viclafargue in #8376
• Improve multi-GPU KMeans validation by @viclafargue in #8350
• Separate sklearn example compatibility and total pass rates by @csadorf in #8409
• Make FFT t-SNE deterministic when random seed is set by @dantegd in #8123
• Fix Dask logistic regression with non-contiguous class labels by @viclafargue in #8422

📖 Documentation

• Modernize the Python developer docs by @csadorf in #7642
• Update docs to use NVIDIA Sphinx theme by @bdice in #8308
• Restructure and clarify cuml.accel documentation by @csadorf in #8367
• Document list and tuple input support by @csadorf in #8381

🚀 New Features

• Add IPython magic for setting cuml.accel log level by @jcrist in #8231
• [FEA] Add host-streaming support to single-GPU KMeans by @tarang-jain in #8248
• Isolation Forest Implementation by @dantegd in #8226

🛠️ Improvements

• Suppress intentional hardcoded version references by @csadorf in #8113
• Forward merge release/26.06 into main by @csadorf in #8138
• Improve devcontainer startup times by @trxcllnt in #8135
• Fix symbol export by @vyasr in #8037
• Fix a few lingering sklearn compat xfails by @jcrist in #8166
• Add YAML based benchmark definitions for cuML regression runs by @dantegd in #7980
• Report xfail details in Python test output by @csadorf in #8140
• Add SECURITY.md by @jameslamb in #8175
• merge release/26.06 into main by @jameslamb in #8184
• Replace RF global quantile sort with deterministic sampled quantiles by @RAMitchell in #8111
• Large (int64 indices) sparse input support by @jcrist in #8196
• RF: narrow CountBin to double for weighted-training prep by @switch527 in #8132
• fix(ci): fix configuration for breaking change notification workflow by @jameslamb in #8199
• TST Make check_estimator coverage opt-out by @betatim in #8176
• Forward merge release/26.06 into main by @csadorf in #8214
• Remove deprecated sample_weight in naive bayes/sample_weights in hinge_loss by @jcrist in #8221
• Remove deprecated probability argument to SVC by @jcrist in #8223
• Bump scikit-learn minimum to 1.6 by @csadorf in #8091
• Avoid host sync in _repr_mimebundle_ when possible by @jcrist in #8232
• Remove deprecated probability argument to LinearSVC by @jcrist in #8224
• Fix type reflection when input is an array-like by @jcrist in #8234
• Update cuml.tsa to use new validation by @jcrist in #8236
• Unify distributed RF quantile samples by @RAMitchell in #8190
• Remove numba-cuda upper bound by @brandon-b-miller in #8189
• refactor: switch to rapids-artifact-name for consistent artifact naming by @gforsyth in #8174
• Update cuML sklearn tags for scikit-learn 1.6 by @csadorf in #8227
• Refactor RF objectives by objective family by @RAMitchell in #8233
• Update multi-gpu estimators to new input validation by @jcrist in #8241
• Prepare RF bins and objectives for weights support by @RAMitchell in #8247
• A few fixups to multi-gpu implementations by @jcrist in #8249
• Remove obsolete scikit-learn 1.5 compatibility paths by @csadorf in #8253
• Remove stale sklearn example xfail by @csadorf in #8259
• Build and test with CUDA 13.3.0 by @bdice in #8246
• Remove make_monotonic and sorted_unique_labels by @jcrist in #8269
• Cleanup cuml.model_selection by @jcrist in #8270
• Fixup CI by @jcrist in #8280
• Limit high-dimensional cosine KDE property cases by @Federicorao in #8273
• Validate random forest n_streams by @fallintoplace in #8263
• Use main shared-workflows branch by @bdice in #8286
• Mark test_logreg_l1[42-csr_array] flaky by @jcrist in #8289
• Use segmented scan for RF row partitioning by @RAMitchell in #8257
• Consolidate sparse utilities by @jcrist in #8288
• Validate LinearSVC n_streams by @fallintoplace in #8282
• Add C++ RF sample weight training support by @RAMitchell in #8258
• Center RF split thresholds across empty quantile bins by @RAMitchell in #8283
• Remove xfail markers for cuml.accel examples by @jinsolp in #8297
• FIX Make __repr__ show parameters set to non-default values by @betatim in #8229
• Avoid hard-coded SHAP test outputs by @viclafargue in #8264
• CI Run cuml.accel tests in the sklearn nightly test runs by @betatim in #8219
• Simplify reflect decorator by @jcrist in #8295
• Remove cuml.internals.input_utils and remaining legacy ingest paths by @jcrist in #8293
• Remove RMM make_any_device_resource helper usage by @bdice in #8197
• skip Cython 3.2.7 by @jameslamb in #8301
• Update CuPy and NumPy dependency specs by @bdice in #8287
• Implement weighted bootstrap sampling for RF by @RAMitchell in #8296
• Add benchmark coverage for missing cuML estimators by @viclafargue in #8271
• Warn in cuml.accel if proxying a package outside a range of tested versions by @jcrist in #8303
• cuml.accel SVM onnx support by @jcrist in #8307
• Thread sample weights through RandomForest Python API by @RAMitchell in #8306
• Add benchmark coverage for missing cuML estimators 2 by @viclafargue in #8325
• Add benchmark coverage for missing cuML estimators 3 by @viclafargue in #8341
• switch from 'cuda-python' to specific components (e.g. 'cuda-bindings') by @jameslamb in #8314
• Handle RF split histogram shared-memory pressure by @RAMitchell in #8323
• Empty commit to trigger new builds due to RAFT 3052 ABI breakage by @divyegala in #8353
• fix(naive_bayes): use established classes_ on later partial_fit calls by @nethum529 in #8316
• Prevent cuVS types from leaking through public headers by @viclafargue in #8342
• Support row-major RandomForest training inputs by @RAMitchell in #8324
• Add regression test for cluster-metric label-permutation invariance (#7199) by @aminehd in #8355
• Prepare RF split counts for distributed training by @RAMitchell in #8348
• Raise TypeError for invalid sparse input types in benchmark datagen by @fallintoplace in #8318
• CI Fix bertopic testing by @betatim in #8298
• Unify reflection system & decorators by @jcrist in #8339
• Add JSON output for benchmark results by @dantegd in #8027
• Increase cython-lint line length by @jcrist in #8374
• A few sklearn compat test fixes by @jcrist in #8366
• Split random forest histogram building and split scoring kernels by @RAMitchell in #8370
• Re-enable RF regressor C++ benchmarks by @RAMitchell in #8371
• Support coercing host arrays in mlfunc/ReflectedAttr by @jcrist in #8373
• Finish migration to new decorators/descriptors by @jcrist in #8378
• Address signedness compiler warning by @chyunsu3 in #8382
• Add RF histogram allreduce hook for distributed split scoring by @RAMitchell in #8377
• Enable building with 26.08 nvForest by @chyunsu3 in #8388
• Remove legacy array and reflection machinery by @jcrist in #8385
• Update developer wiki for new machinery by @jcrist in #8386
• Empty commit to trigger a build by @bdice in #8399
• [FEA] Skip Device Concat for KmeansMG by @tarang-jain in #8084

New Contributors

• @Federicorao made their first contribution in #8273
• @nethum529 made their first contribution in #8316
• @Husnain987 made their first contribution in #8292
• @aminehd made their first contribution in #8355

Full Changelog: v26.08.00a...v26.08.00