2.1.2 Patch Release

The 2.1.2 patch release makes the following bug fixes:

• Clean up and modernize release-artifacts.py (#10818)
• Fix ellpack categorical feature with missing values. (#10906)
• Fix unbiased ltr with training continuation. (#10908)
• Fix potential race in feature constraint. (#10719)
• Fix boolean array for arrow-backed DF. (#10527)
• Ensure that pip check does not fail due to a bad platform tag (#10755)
• Check cub errors (#10721)
• Limit the maximum number of threads. (#10872)
• Fixes for large size clusters. (#10880)
• POSIX compliant poll.h and mmap (#10767)

Additional artifacts:

You can verify the downloaded packages by running the following command on your Unix shell:

echo "<hash> <artifact>" | shasum -a 256 --check

a84fc7d9846c24659a2ad16788a7eefa9640b19eea9bbc65f30e0a9d53c52453  xgboost-2.1.2.tar.gz
999eff38533ea79ab3a1f0da524c54f6d0abd2ef220b6dbb9ba1331703e898bc  xgboost_r_gpu_linux_2.1.2.tar.gz

Experimental binary packages for R with CUDA enabled

• xgboost_r_gpu_linux_2.1.2.tar.gz: Download

Source tarball

• xgboost.tar.gz: Download

2.1.1 Patch Release

The 2.1.1 patch release make the following bug fixes:

• [Dask] Disable broadcast in the scatter call so that predict function won't hang (#10632) by @trivialfis
• [Dask] Handle empty partitions correctly (#10559) by @trivialfis
• Fix federated learning for the encrypted GRPC backend (#10503) by @trivialfis
• Fix a race condition in column splitter (#10572) by @trivialfis
• Gracefully handle cases where system files like /sys/fs/cgroup/cpu.max are not readable by the user (#10623) by @trivialfis
• Fix build and C++ tests for FreeBSD (#10480) by @hcho3, @trivialfis
• Clarify the requirement Pandas 1.2+ (#10476) by @hcho3
• More robust endianness detection in R package build (#10642) by @jakirkham, @hcho3

In addition, it contains several enhancements:

• Publish JVM packages targeting Linux ARM64 (#10487) by @hcho3
• Publish a CPU-only wheel under name xgboost-cpu (#10603) by @hcho3
• Support building with CUDA Toolkit 12.5 and latest CCCL (#10624, #10633, #10574) by @hcho3, @trivialfis, @jakirkham

Full Changelog: v2.1.0...v2.1.1

Additional artifacts:

You can verify the downloaded packages by running the following command on your Unix shell:

echo "<hash> <artifact>" | shasum -a 256 --check

eddbc5200b7c5210f2b8974b9d2a0328a30753416bfb81fdaf5040f4f7abb222  xgboost-2.1.1.tar.gz
3ba5a6e0c609bd5cc0a667d83c57457c06778bece50863e58c8bc1b4eb415fc6  xgboost_r_gpu_linux_2.1.1.tar.gz

Experimental binary packages for R with CUDA enabled

• xgboost_r_gpu_linux_2.1.1.tar.gz: Download

Source tarball

• xgboost.tar.gz: Download

Release 2.1.0 stable

2.1.0 (2024 Jun 20)

We are thrilled to announce the XGBoost 2.1 release. This note will start by summarizing some general changes and then highlighting specific package updates. As we are working on a new R interface, this release will not include the R package. We'll update the R package as soon as it's ready. Stay tuned!

Networking Improvements

An important ongoing work for XGBoost, which we've been collaborating on, is to support resilience for improved scaling and federated learning on various platforms. The existing networking library in XGBoost, adopted from the RABIT project, can no longer meet the feature demand. We've revamped the RABIT module in this release to pave the way for future development. The choice of using an in-house version instead of an existing library is due to the active development status with frequent new feature requests like loading extra plugins for federated learning. The new implementation features:

• Both CPU and GPU communication (based on NCCL).
• A reusable tracker for both the Python package and JVM packages. With the new release, the JVM packages no longer require Python as a runtime dependency.
• Supports federated communication patterns for both CPU and GPU.
• Supports timeout. The high-level interface parameter is currently hard-coded to 30 minutes, which we plan to improve.
• Supports significantly more data types.
• Supports thread-based workers.
• Improved handling for worker errors, including better error messages when one of the peers dies during training.
• Work with IPv6. Currently, this is only supported by the dask interface.
• Built-in support for various operations like broadcast, allgatherV, allreduce, etc.

Related PRs (#9597, #9576, #9523, #9524, #9593, #9596, #9661, #10319, #10152, #10125, #10332, #10306, #10208, #10203, #10199, #9784, #9777, #9773, #9772, #9759, #9745, #9695, #9738, #9732, #9726, #9688, #9681, #9679, #9659, #9650, #9644, #9649, #9917, #9990, #10313, #10315, #10112, #9531, #10075, #9805, #10198, #10414).

The existing option of using MPI in RABIT is removed in the release. (#9525)

NCCL is now fetched from PyPI.

In the previous version, XGBoost statically linked NCCL, which significantly increased the binary size and led to hitting the PyPI repository limit. With the new release, we have made a significant improvement. The new release can now dynamically load NCCL from an external source, reducing the binary size. For the PyPI package, the nvidia-nccl-cu12 package will be fetched during installation. With more downstream packages reusing NCCL, we expect the user environments to be slimmer in the future as well. (#9796, #9804, #10447)

Parts of the Python package now require glibc 2.28+

Starting from 2.1.0, XGBoost Python package will be distributed in two variants:

• manylinux_2_28: for recent Linux distros with glibc 2.28 or newer. This variant comes with all features enabled.
• manylinux2014: for old Linux distros with glibc older than 2.28. This variant does not support GPU algorithms or federated learning.

The pip package manager will automatically choose the correct variant depending on your system.

Starting from May 31, 2025, we will stop distributing the manylinux2014 variant and exclusively distribute the manylinux_2_28 variant. We made this decision so that our CI/CD pipeline won't have depend on software components that reached end-of-life (such as CentOS 7). We strongly encourage everyone to migrate to recent Linux distros in order to use future versions of XGBoost.

Note. If you want to use GPU algorithms or federated learning on an older Linux distro, you have two alternatives:

1. Upgrade to a recent Linux distro with glibc 2.28+. OR
2. Build XGBoost from the source.

Multi-output

We continue the work on multi-target and vector leaf in this release:

• Revise the support for custom objectives with a new API, XGBoosterTrainOneIter. This new function supports strided matrices and CUDA inputs. In addition, custom objectives now return the correct shape for prediction. (#9508)
• The hinge objective now supports multi-target regression (#9850)
• Fix the gain calculation with vector leaf (#9978)
• Support graphviz plot for multi-target tree. (#10093)
• Fix multi-output with alternating strategies. (#9933)

Please note that the feature is still in progress and not suitable for production use.

Federated Learning

Progress has been made on federated learning with improved support for column-split, including the following updates:

• Column split work for both CPU and GPU. In addition, categorical data is now compatible with column split. (#9562, #9609, #9611, #9628, #9539, #9578, #9685, #9623, #9613, #9511, #9384, #9595)
• The use of UBJson to serialize split entries for column split has been implemented, aiding vector-leaf with column-based data split. (#10059, #10055, #9702)
• Documentation and small fixes. (#9610, #9552, #9614, #9867)

Ongoing work for SYCL support.

XGBoost is developing a SYCL plugin for SYCL devices, starting with the hist tree method. (#10216, #9800, #10311, #9691, #10269, #10251, #10222, #10174, #10080, #10057, #10011, #10138, #10119, #10045, #9876, #9846, #9682) XGBoost now supports launchable inference on SYCL devices, and work on adding SYCL support for training is ongoing.

Looking ahead, we plan to complete the training in the coming releases and then focus on improving test coverage for SYCL, particularly for Python tests.

Optimizations

• Implement column sampler in CUDA for GPU-based tree methods. This helps us get faster training time when column sampling is employed (#9785)
• CMake LTO and CUDA arch (#9677)
• Small optimization to external memory with a thread pool. This reduces the number of threads launched during iteration. (#9605, #10288, #10374)

Deprecation and breaking changes

Package-specific breaking changes are outlined in respective sections. Here we list general breaking changes in this release:

• The command line interface is deprecated due to the increasing complexity of the machine learning ecosystem. Building a machine learning model using a command shell is no longer feasible and could mislead newcomers. (#9485)
• Universal binary JSON is now the default format for saving models (#9947, #9958, #9954, #9955). See #7547 for more info.
• The XGBoosterGetModelRaw is now removed after deprecation in 1.6. (#9617)
• Drop support for loading remote files. Users are encouraged to use dedicated libraries to fetch remote content. (#9504)
• Remove the dense libsvm parser plugin. This plugin is never tested or documented (#9799)
• XGDMatrixSetDenseInfo and XGDMatrixSetUIntInfo are now deprecated. Use the array interface based alternatives instead.

Features

This section lists some new features that are general to all language bindings. For package-specific changes, please visit respective sections.

• Adopt a new XGBoost logo (#10270)
• Now supports dataframe data format in native XGBoost. This improvement enhances performance and reduces memory usage when working with dataframe-based structures such as pandas, arrow, and R dataframe. (#9828, #9616, #9905)
• Change default metric for gamma regression to deviance. (#9757)
• Normalization for learning to rank is now optional with the introduction of the new lambdarank_normalization parameter. (#10094)
• Contribution prediction with QuantileDMatrix on CPU. (#10043)
• XGBoost on macos no longer bundles OpenMP runtime. Users can install the latest runtime from their dependency manager of choice. (#10440). Along with which, JVM packages on MacoOS are now built with OpenMP support (#10449).

Bug fixes

• Fix training with categorical data from external memory. (#10433)
• Fix compilation with CTK-12. (#10123)
• Fix inconsistent runtime library on Windows. (#10404)
• Fix default metric configuration. (#9575)
• Fix feature names with special characters. (#9923)
• Fix global configuration for external memory training. (#10173)
• Disable column sample by node for the exact tree method. (#10083)
• Fix the FieldEntry constructor specialization syntax error (#9980)
• Fix pairwise objective with NDCG metric along with custom gain. (#10100)
• Fix the default value for lambdarank_pair_method. (#10098)
• Fix UBJSON with boolean values. No existing code is affected by this fix. (#10054)
• Be more lenient on floating point errors for AUC. This prevents the AUC > 1.0 error. (#10264)
• Check support status for categorical features. This prevents gblinear from treating categorical features as numerical. (#9946)

Document

Here is a list of documentation changes not specific to any XGBoost package.

• A new coarse map for XGBoost features to assist development. (#10310)
• New language binding consistency guideline. (#9755, #9866)
• Fixes, cleanups, small updates (#9501, #9988, #10023, #10013, #10143, #9904, #10179, #9781, #10340, #9658, #10182, #9822)
• Update document for parameters (#9900)
• Brief introduction to base_score. (#9882)
• Mention data consistency for categorical features. (#9678)

Python package

• Dask
  Other than the changes in networking, we have some optimizations and document updates in dask:

• Filter models on workers instead of clients; this prevents an OOM error on the client machine. (#9518)
• Users are now encouraged to use from xgboost import dask instead of import xgboost.dask to avoid drawing in unnecessary dependencies for non-dask users. (#9742)
• Add seed to demos. (#10009)
• New document for using dask XGBoost with k8s. (#10271)
• Workaround potentially unaligned pointer from an empty partition. (#10418)
• Workaround a ...

Release candidate of version 2.1.0

See #10356 for details.

2.0.3 Patch Release

The 2.0.3 patch release make the following bug fixes:

• [backport][sklearn] Fix loading model attributes. (#9808) by @trivialfis in #9880
• [backport][py] Use the first found native library. (#9860) by @trivialfis in #9879
• [backport] [CI] Upload libxgboost4j.dylib (M1) to S3 bucket by @hcho3 in #9887
• [jvm-packages] Fix POM for xgboost-jvm metapackage by @hcho3 in #9893 #9897

Full Changelog: v2.0.2...v2.0.3

Additional artifacts:

You can verify the downloaded packages by running the following command on your Unix shell:

echo "<hash> <artifact>" | shasum -a 256 --check

7c4bd1cf6162d335fd20a8168a54dd11508342f82fbf381a80c02ac57be0bce4  xgboost-2.0.3.tar.gz
d0c3499504133a8ea0043da2974c51cc71aae792f0719080bc227d7add8fb881  xgboost_r_gpu_win64_2.0.3.tar.gz
ee47da5b21231965b1f054d191a5418543377f4ba0d0615a593a6f99d1832ca1  xgboost_r_gpu_linux_2.0.3.tar.gz

Experimental binary packages for R with CUDA enabled

• xgboost_r_gpu_linux_2.0.3.tar.gz: Download
• xgboost_r_gpu_win64_2.0.3.tar.gz: Download

2.0.2 Patch Release

The 2.0.2 patch releases make the following bug fixes:

• [jvm-packages] Add Scala version suffix to xgboost-jvm package (#9776). The JVM packages had incorrect metadata, and the 2.0.2 patch version fixes the metadata.
• [backport] Fix using categorical data with the ranker. (#9753)

2.0.1 Patch Release

This is a patch release for bug fixes.

Bug fixes

• Support pandas 2.1.0. by @trivialfis in #9655
• Fix default metric configuration. by @trivialfis in #9590
• [R] Fix method name. by @trivialfis in #9592
• Use array interface for testing NumPy arrays. by @trivialfis in #9635
• Workaround Apple clang issue. by @trivialfis in #9636
• Add support for cgroupv2. by @trivialfis in #9656
• Fix build for GCC 8.x by @hcho3 in #9670
• [pyspark] Support stage-level scheduling by @wbo4958 in #9686
• Fix build for AppleClang 11 by @hcho3 in #9684
• Fix libpath logic for Windows by @hcho3 in #9687, #9711
• Remove hard dependency on libjvm by @hcho3 in #9705

In addition, this is the first release where the JVM package is distributed with native support for Apple Silicon.

Additional artifacts:

You can verify the downloaded packages by running the following command on your Unix shell:

echo "<hash> <artifact>" | shasum -a 256 --check

529e9d0f88c2a7abae833f05b7d1e7e7ce01de20481ea60f6ebb6eb7fc96ba69  xgboost.tar.gz
25342c91e7cda98b1362b70282b286c2e4f3e996b518fb590c1303f53f39f188  xgboost_r_gpu_win64_2.0.1.tar.gz
3d8cde1160ab135c393b8092ce0475709dff318024022b735a253d968f9711b3  xgboost_r_gpu_linux_2.0.1.tar.gz

Experimental binary packages for R with CUDA enabled

• xgboost_r_gpu_linux_2.0.1.tar.gz: Download
• xgboost_r_gpu_win64_2.0.1.tar.gz: Download

Source tarball

• xgboost.tar.gz: Download

Release 2.0.0 stable

2.0.0 (2023 Sep 12)

We are excited to announce the release of XGBoost 2.0. This note will begin by covering some overall changes and then highlight specific updates to the package.

Initial work on multi-target trees with vector-leaf outputs

We have been working on vector-leaf tree models for multi-target regression, multi-label classification, and multi-class classification in version 2.0. Previously, XGBoost would build a separate model for each target. However, with this new feature that's still being developed, XGBoost can build one tree for all targets. The feature has multiple benefits and trade-offs compared to the existing approach. It can help prevent overfitting, produce smaller models, and build trees that consider the correlation between targets. In addition, users can combine vector leaf and scalar leaf trees during a training session using a callback. Please note that the feature is still a working in progress, and many parts are not yet available. See #9043 for the current status. Related PRs: (#8538, #8697, #8902, #8884, #8895, #8898, #8612, #8652, #8698, #8908, #8928, #8968, #8616, #8922, #8890, #8872, #8889, #9509) Please note that, only the hist (default) tree method on CPU can be used for building vector leaf trees at the moment.

New device parameter.

A new device parameter is set to replace the existing gpu_id, gpu_hist, gpu_predictor, cpu_predictor, gpu_coord_descent, and the PySpark specific parameter use_gpu. Onward, users need only the device parameter to select which device to run along with the ordinal of the device. For more information, please see our document page (https://xgboost.readthedocs.io/en/stable/parameter.html#general-parameters) . For example, with device="cuda", tree_method="hist", XGBoost will run the hist tree method on GPU. (#9363, #8528, #8604, #9354, #9274, #9243, #8896, #9129, #9362, #9402, #9385, #9398, #9390, #9386, #9412, #9507, #9536). The old behavior of gpu_hist is preserved but deprecated. In addition, the predictor parameter is removed.

hist is now the default tree method

Starting from 2.0, the hist tree method will be the default. In previous versions, XGBoost chooses approx or exact depending on the input data and training environment. The new default can help XGBoost train models more efficiently and consistently. (#9320, #9353)

GPU-based approx tree method

There's initial support for using the approx tree method on GPU. The performance of the approx is not yet well optimized but is feature complete except for the JVM packages. It can be accessed through the use of the parameter combination device="cuda", tree_method="approx". (#9414, #9399, #9478). Please note that the Scala-based Spark interface is not yet supported.

Optimize and bound the size of the histogram on CPU, to control memory footprint

XGBoost has a new parameter max_cached_hist_node for users to limit the CPU cache size for histograms. It can help prevent XGBoost from caching histograms too aggressively. Without the cache, performance is likely to decrease. However, the size of the cache grows exponentially with the depth of the tree. The limit can be crucial when growing deep trees. In most cases, users need not configure this parameter as it does not affect the model's accuracy. (#9455, #9441, #9440, #9427, #9400).

Along with the cache limit, XGBoost also reduces the memory usage of the hist and approx tree method on distributed systems by cutting the size of the cache by half. (#9433)

Improved external memory support

There is some exciting development around external memory support in XGBoost. It's still an experimental feature, but the performance has been significantly improved with the default hist tree method. We replaced the old file IO logic with memory map. In addition to performance, we have reduced CPU memory usage and added extensive documentation. Beginning from 2.0.0, we encourage users to try it with the hist tree method when the memory saving by QuantileDMatrix is not sufficient. (#9361, #9317, #9282, #9315, #8457)

Learning to rank

We created a brand-new implementation for the learning-to-rank task. With the latest version, XGBoost gained a set of new features for ranking task including:

• A new parameter lambdarank_pair_method for choosing the pair construction strategy.
• A new parameter lambdarank_num_pair_per_sample for controlling the number of samples for each group.
• An experimental implementation of unbiased learning-to-rank, which can be accessed using the lambdarank_unbiased parameter.
• Support for custom gain function with NDCG using the ndcg_exp_gain parameter.
• Deterministic GPU computation for all objectives and metrics.
• NDCG is now the default objective function.
• Improved performance of metrics using caches.
• Support scikit-learn utilities for XGBRanker.
• Extensive documentation on how learning-to-rank works with XGBoost.

For more information, please see the tutorial. Related PRs: (#8771, #8692, #8783, #8789, #8790, #8859, #8887, #8893, #8906, #8931, #9075, #9015, #9381, #9336, #8822, #9222, #8984, #8785, #8786, #8768)

Automatically estimated intercept

In the previous version, base_score was a constant that could be set as a training parameter. In the new version, XGBoost can automatically estimate this parameter based on input labels for optimal accuracy. (#8539, #8498, #8272, #8793, #8607)

Quantile regression

The XGBoost algorithm now supports quantile regression, which involves minimizing the quantile loss (also called "pinball loss"). Furthermore, XGBoost allows for training with multiple target quantiles simultaneously with one tree per quantile. (#8775, #8761, #8760, #8758, #8750)

L1 and Quantile regression now supports learning rate

Both objectives use adaptive trees due to the lack of proper Hessian values. In the new version, XGBoost can scale the leaf value with the learning rate accordingly. (#8866)

Export cut value

Using the Python or the C package, users can export the quantile values (not to be confused with quantile regression) used for the hist tree method. (#9356)

column-based split and federated learning

We made progress on column-based split for federated learning. In 2.0, both approx, hist, and hist with vector leaf can work with column-based data split, along with support for vertical federated learning. Work on GPU support is still on-going, stay tuned. (#8576, #8468, #8442, #8847, #8811, #8985, #8623, #8568, #8828, #8932, #9081, #9102, #9103, #9124, #9120, #9367, #9370, #9343, #9171, #9346, #9270, #9244, #8494, #8434, #8742, #8804, #8710, #8676, #9020, #9002, #9058, #9037, #9018, #9295, #9006, #9300, #8765, #9365, #9060)

PySpark

After the initial introduction of the PySpark interface, it has gained some new features and optimizations in 2.0.

• GPU-based prediction. (#9292, #9542)
• Optimization for data initialization by avoiding the stack operation. (#9088)
• Support predict feature contribution. (#8633)
• Python typing support. (#9156, #9172, #9079, #8375)
• use_gpu is deprecated. The device parameter is preferred.
• Update eval_metric validation to support list of strings (#8826)
• Improved logs for training (#9449)
• Maintenance, including refactoring and document updates (#8324, #8465, #8605, #9202, #9460, #9302, #8385, #8630, #8525, #8496)
• Fix for GPU setup. (#9495)

Other General New Features

Here's a list of new features that don't have their own section and yet are general to all language bindings.

• Use array interface for CSC matrix. This helps XGBoost to use a consistent number of threads and align the interface of the CSC matrix with other interfaces. In addition, memory usage is likely to decrease with CSC input thanks to on-the-fly type conversion. (#8672)
• CUDA compute 90 is now part of the default build.. (#9397)

Other General Optimization

These optimizations are general to all language bindings. For language-specific optimization, please visit the corresponding sections.

• Performance for input with array_interface on CPU (like numpy) is significantly improved. (#9090)
• Some optimization with CUDA for data initialization. (#9199, #9209, #9144)
• Use the latest thrust policy to prevent synchronizing GPU devices. (#9212)
• XGBoost now uses a per-thread CUDA stream, which prevents synchronization with other streams. (#9416, #9396, #9413)

Notable breaking change

Other than the aforementioned change with the device parameter, here's a list of breaking changes affecting all packages.

• Users must specify the format for text input (#9077). However, we suggest using third-party data structures such as numpy.ndarray instead of relying on text inputs. See #9472 for more info.

Notable bug fixes

Some noteworthy bug fixes that are not related to specific language bindings are listed in this section.

• Some language environments use a different thread to perform garbage collection, which breaks the thread-local cache used in XGBoost. XGBoost 2.0 implements a new thread-safe cache using a light weight lock to replace the thread-local cache. (#8851)
• Fix model IO by clearing the prediction cache. (#8904)
• inf is checked during data construction. (#8911)
• Preserve order of saved updaters configuration. Usually, this is not an issue unless the updater parameter is used instead of the tree_method parameter (#9355)
• Fix GPU memory allocation issue with categorical splits. (#9529)
• Handle escape sequence like \t\n in feature names for JSON model dump. (#9474)
• Normalize file path for model IO and text input. This handles short paths on Windows and paths that contain ~ on Unix (#9463). In addition, all...

Release candidate of version 2.0.0

Roadmap: https://github.com/dmlc/xgboost/projects/2
Release note: #9484
Release status: #9497

1.7.6 Patch Release

This is a patch release for bug fixes. The CRAN package for the R binding is kept at 1.7.5.

Bug Fixes

• Fix distributed training with mixed dense and sparse partitions. (#9272)
• Fix monotone constraints on CPU with large trees. (#9122)
• [spark] Make the spark model have the same UID as its estimator (#9022)
• Optimize prediction with QuantileDMatrix. (#9096)

Document

• Improve doxygen (#8959)
• Update the cuDF pip index URL. (#9106)

Maintenance

• Fix tests with pandas 2.0. (#9014)

Additional artifacts:

You can verify the downloaded packages by running the following command on your Unix shell:

echo "<hash> <artifact>" | shasum -a 256 --check

0a54300dd274b98b7f039acffa006bec4875dace041fd9288422306fe7c379ca  xgboost.tar.gz
990fb3c54be7ce53365389f2eb82ce3c1f2e78735b4605ddd2ddb0d47a15d3c3  xgboost_r_gpu_linux_1.7.6.tar.gz
a48fc64bce774bb76eddade6dc6df1d4fc25199a0c17dc66cdfa50cedd3282ad  xgboost_r_gpu_win64_1.7.6.tar.gz

Experimental binary packages for R with CUDA enabled

• xgboost_r_gpu_linux_1.7.6.tar.gz: Download
• xgboost_r_gpu_win64_1.7.6.tar.gz: Download

Source tarball
Link in GitHub release assets