ContextQMD
Back to Xgboost

3.4.0 (Aug 4, 2026)

doc/changes/v3.4.0.rst

3.4.05.7 KB
Original Source

################### 3.4.0 (Aug 4, 2026) ###################

XGBoost 3.4.0 completes the vector-leaf implementation for the hist tree method, revises the quantile regression and mean absolute error (MAE) objectives, removes column-split support, and fixes issues across multiple components.

Vector Leaf and Multi-output Models

The vector-leaf implementation for the hist tree method is now feature-complete and classified as experimental rather than work in progress. Key changes include:

  • Categorical features on CPU and GPU, DART, monotonic and interaction constraints, and min_child_weight. (:pr:12305, :pr:12299, :pr:12276, :pr:12340, :pr:12341, :pr:12294, :pr:12325, :pr:12296)
  • Distributed training and Dask multi-label estimators, including AUC and SHAP support, plus tree dumping and DataFrame conversion. (:pr:12292, :pr:12369, :pr:12313, :pr:12314, :pr:12293)
  • Batched split application, reduced-gradient sampling, shared histogram building and split evaluation, additional validation, and external-memory fixes. (:pr:12365, :pr:12321, :pr:12397, :pr:12387, :pr:12336, :pr:12330, :pr:12320, :pr:12312)

Quantile Regression and Absolute Error

Quantile regression and mean absolute error now use smooth approximations instead of line searches to determine leaf values. (:pr:12391, :pr:12386, :pr:12373, :pr:12346)

Breaking Changes and Deprecations

  • Column-split support has been removed. (:pr:12333, :pr:12363, :pr:12354)
  • The federated learning plugin is no longer included in Python binary wheels. (:pr:12376)
  • The random-forest wrappers are deprecated. Use num_parallel_tree directly for random-forest models. (:pr:12324, :pr:12342)
  • The deprecated XGDMatrixCreateFromFile function has been removed. (:pr:12297)
  • The CUDA asynchronous memory pool is now recommended for external-memory training. (:pr:12337)

Build and Platform

  • The default xgboost <https://pypi.org/project/xgboost/>__ binaries for the Python and JVM packages are now built with CUDA 13.3 on Linux and Windows. Python users who need CUDA 12.9 can install xgboost-cu12 <https://pypi.org/project/xgboost-cu12/>__ from PyPI or build from source. (:pr:12384, :pr:12394)
  • Both the xgboost <https://pypi.org/project/xgboost/>__ and xgboost-cpu <https://pypi.org/project/xgboost-cpu/>__ packages support Windows on AArch64. (:pr:12394)

Optimizations

  • Reduce allocations in the JSON text parser. (:pr:12317)
  • Skip per-tree depth computation for single-row prediction. (:pr:12307)
  • Avoid rebuilding the regular expression when parsing the device parameter. (:pr:12306)
  • Extend the depth bucket used by GPU Quadrature TreeSHAP. (:pr:12290)

Bug Fixes

The following fixes affect all interfaces. Interface-specific fixes are listed in the corresponding package sections.

  • Fix an issue where min_child_weight could produce an empty root node. Also, this release fixes its interaction with max_delta_step. (:pr:12322, :pr:12296)
  • Preserve tiny positive Hessians during GPU quantization. (:pr:12266)
  • Fix compatibility with RMM 26.08. (:pr:12316, :pr:12269)
  • Fix learning-to-rank pair sampling when continuing training. (:pr:12332)
  • Fix GPU SHAP on CUDA SM120. (:pr:12368)
  • Fix the recoder's handling of UTF-8 data and pandas nullable categorical indexes, and add overflow checks. (:pr:12371)

Python Package

  • Migrate Python package builds to scikit-build-core and include license files in both installed packages and wheel metadata. (:pr:12219, :pr:12349, :pr:12280)
  • :py:meth:~xgboost.Booster.trees_to_dataframe now uses pandas NA for unavailable vector-leaf values instead of a mix of np.nan and None. (:pr:12293)
  • Correct scikit-learn input tags and raise ValueError when qid is omitted from :py:meth:~xgboost.XGBRanker.score. (:pr:12383, :pr:12335)
  • Pass Graphviz keyword arguments through :py:func:~xgboost.plotting.to_graphviz. (:pr:12359)
  • Fix shape reporting for empty array-interface results and SciPy CSC input handling. (:pr:12381, :pr:12378)
  • Fix :py:func:~xgboost.collective.allreduce always returning a flattened one-dimensional array. (:pr:12377)
  • Code cleanup. (:pr:12331)

R Package

  • Fix early_stopping_rounds handling in xgboost(). (:pr:12370)
  • Remove an obsolete 32-bit Windows thread_local workaround. (:pr:12352)
  • Test and lint fixes. (:pr:12309, :pr:12259)

JVM Packages

  • Fix potential resource leaks in native error paths and make floating-point serialization locale-independent. (:pr:12286, :pr:12235)
  • Update the Maven Central publishing plugin and release tooling, including support for local testing. (:pr:12255, :pr:12254, :pr:12253)

Documentation and Demos

  • Add the second-place solution from the Tabular Playground Series. (:pr:12392)
  • Fix typos and update external links. (:pr:12400, :pr:12401, :pr:12357, :pr:12358, :pr:12328)

CI and Maintenance

  • Update GitHub Actions. (:pr:12375, :pr:12338, :pr:12300, :pr:12263, :pr:12274)
  • Add a stale pull request cleanup workflow. (:pr:12318)
  • Clean up the NCCL discovery script. (:pr:12262)
  • Refactor and clean up code. (:pr:12345, :pr:12275, :pr:12385, :pr:12271)
  • Release maintenance. (:pr:12252)