1.2.9

[andrey-khropov]

⚠️ There are no JVM artifacts for this release due to issues with publishing.: They are available in the next release 1.2.10.

Major changes

• [Python-package] Add polars input data support. Support polars input #2524.

  Polars data structures are supported for features, labels and auxiliary data like weight, timestamp etc.

New features

• [R-package] Make 'predict' an S3 method (R) Make 'predict' an S3 method #1657. Thanks to @david-cortes.
• Add RMSPE metric and loss (both as are CPU-only for now) RMSPE Metric #1767. Thanks to @ivan339339.
• [C/C++ applier] New function LoadFullModelZeroCopy for mmap New C API function LoadFullModelZeroCopy for mmap #2893. Thanks to @gakoshin.

Improvements

• Remove the limit of 128 threads when loading data. CatBoost cannot use more than 128 threads when loading data #3027

Speedups

• Optimize Lossguide grow policy on CPU Optimize Lossguide #2883. Approximate speedup is 1.4x. Thanks to @Levachev.
• [Python-package] Support non-float32 numpy numeric types in multithreaded native features data initialization. Pool does not use multiple cores during initialisation  #1558, Optimize catboost pool initialization for np.float64 case #2847
• [Python-package] Avoid possible repeated reparsing of estimator parameters to canonical forms

Python package

• Support Python 3.14 Support Python 3.14 #2943
• pyproject.toml is now PEP-517 compliant.
• Estimators: Add __sklearn_tags__ method to be compatible with scikit-learn >= 1.8.x. AttributeError: The following error was raised: 'CatBoostRegressor' object has no attribute '__sklearn_tags__' with scikit-learn 1.8.x #2955
• Estimators: Add __repr__ method with a meaningful description expected by scikit-learn CatBoostRegressor produces meaningless repr #2307. Thanks to @besteady.
• Adapt to the removal of dry_run parameters in setuptools 81.0. pypa/setuptools#4872
• Set upper version bounds for important dependencies to avoid breaking changes

Rust package

• Implement Sync for rust Model struct Sync is not implemented for model struct in rust #2689. Thanks to @alexeysofin.
• Support Windows Rust 2024 bindings for Microsoft Windows #2842
• Fix build on Linux aarch64 Errors when building on aarch64 target with Rust #2890. Thanks to @joelchen.

Build & testing

• [Windows] Support building with MSVC toolsets 14.4*. Support working on CatBoost with Visual Studio 2022 #2302
• [GPU] Switch from hardcoded gencode specifications in multiple CMakeLists.txt files to the standard CMake variable CMAKE_CUDA_ARCHITECTURES, although the default value is non-standard and specified in cuda.cmake. Respect CMAKE_CUDA_ARCHITECTURES #2540
• [macOS] Support Apple Clang 17. Apple clang version 17.0.0 (clang-1700.0.13.3) build failed #2860
• [Python-package][Windows] Fix python package installation from a source distribution. python package installation from source distribution is broken on Windows #3024
• [Python-package] wheel build dependency no longer required

Bugfixes

• [Performance][Windows] __SSE__ compiler flag was not enabled for Windows builds with MSVC compiler. This affected code that relied on this flag including some operations used during training and quantization during model inference. It is important to note that the compiler itself was configured for SSE support and could still apply automatic SSE optimizations.
• [Python-package] carry.py: fix _uplift_by_name. Possible bug in carry.py: _uplift_by_name assigned to _uplift_by_index #2861
• [Python-package] CatBoostError was missing from __all__ in catboost package. 'CatBoostError' is missing from __init__.py #2862
• [Python-package] log_cout was used instead of log_cerr by mistake. Possible bug in core.py, _CustomLoggersStack.push() ignores log_cerr #2863
• [Python-package] Don't fail when all features are embeddings with the same dimension. Bug: Pool's embedding_features raises exception with same-sized vectors #2875
• [Python-package] get_params: deep parameter meaning was inconsistent with scikit-learn expectations. CatBoost estimators: get_params: deep parameter meaning is inconsistent with scikit-learn expectations #2991
• [Python-package] Estimators' _get_tags: Add missing tags. _get_tags lacks necessary members for scikit-learn < 1.6.0 #3008
• [Python-package] Estimators' _get_tags returned incorrect values for several tags. _get_tags. Incorrectly set tags #3009
• [Python-package] Incorrect values were silently accepted in timestamp parameters. [python-package] Incorrect values are accepted in timestamp parameters #3019
• [CLI] fix eval result output for MultiRMSE
• [GPU] Fix devices parameter parsing. Parsing was non-robust: in case of non-numbers specified it defaulted to 0 and device ids outside of the available range were silently ignored.
• [C/C++ applier] Fix a race condition in error messages reported by GetErrorString in multithreaded programs. It is now thread-local.