RSN 62 - Removing rapids-xgboost conda package in RAPIDS Release v26.10

Author	RAPIDS TPM
Status	In Progress
Topic	Platform Support Change
RAPIDS Version	All RAPIDS Versions
Created	July 30, 2026
Updated	September 30, 2026

Overview

RAPIDS v26.08 is the final release that publishes RAPIDS-specific XGBoost conda packages, including rapids-xgboost. Starting with RAPIDS v26.10, the rapids metapackage uses XGBoost from conda-forge instead.

Historical packages have also changed to allow users to install XGBoost from conda-forge, including when using strict channel priority.

Impact

Conda channels

Historical xgboost, libxgboost, py-xgboost, r-xgboost, xgboost-proc, and rapids-xgboost packages have moved out of the rapidsai channel’s default main label to rapidsai/label/legacy-xgboost. These packages remain available under that label; they have not been deleted.

With strict channel priority, keeping any xgboost packages on the higher-priority rapidsai channel would prevent conda from selecting newer versions from conda-forge. Moving the historical packages removes that restriction.

Historical RAPIDS releases

The rapids metapackages for RAPIDS v25.12 through v26.08 have been republished with unchanged package versions. These builds depend directly on xgboost instead of rapids-xgboost, preserving the XGBoost version constraints but removing the rapidsai_* build requirement. With the default channels, these installations now use XGBoost from conda-forge and no longer include RMM integration in XGBoost.

The patched metapackages display an install-time warning with a restoration command pinned to the installed RAPIDS release. For example, rapids=26.08 displays:

The xgboost in the rapids metapackage no longer supports RMM integration.
It can be restored by installing:
  conda install -c rapidsai/label/legacy-xgboost "rapids=26.08" "xgboost=*=rapidsai_*"
See https://docs.nvidia.com/datascience/notices/rsn0062/ for more information.

The warning omits -c rapidsai; include it when running that command to resolve the explicitly requested rapids metapackage.

The original, unpatched rapids metapackages for these releases will be moved to rapidsai/label/legacy-xgboost. Installation specifications pinned to those builds should be updated to use the patched builds.

The rapids metapackages older than v25.12 have not been patched. Recreating those environments, or any environment that explicitly requires rapids-xgboost, requires the legacy-xgboost label as described below.

Migration guidance

Use standard XGBoost packages

Existing users should migrate workloads away from rapids-xgboost. Use the xgboost package from conda-forge, or the wheels from the Python Package Index PyPI. These packages do not include RAPIDS RMM integration.

In conda installation commands and environment files, replace explicit rapids-xgboost or rapidsai::xgboost requirements with conda-forge::xgboost, using a version compatible with your workload.

Unlike the CUDA-only RAPIDS packages, conda-forge provides separate CPU and CUDA builds of XGBoost. When creating the environment, select the desired CUDA Toolkit version with the cuda-version package. Then explicitly request a CUDA build to ensure GPU support:

conda install -c conda-forge "xgboost=*=cuda*"

Unlike the RAPIDS packages, conda-forge’s CUDA builds require conda to detect an NVIDIA driver at installation time. Run conda info and look for __cuda in the virtual packages section. If __cuda is missing or listed as __cuda=0=0, conda has not detected the driver, possibly because it is not installed.

If driver detection is incorrect, or you are creating an environment on a machine without a driver, set CONDA_OVERRIDE_CUDA to the <MAJOR>.<MINOR> version of the latest CUDA Toolkit supported by the driver on the machine where the environment will run. This overrides conda’s detection; it does not install a driver or remove the need for one to run GPU workloads.

Restore legacy RMM-enabled XGBoost

To restore RMM-enabled XGBoost in an existing RAPIDS environment, run:

conda install -c rapidsai/label/legacy-xgboost "xgboost=*=rapidsai_*"

To recreate an environment that depends on historical rapids-xgboost packages, add -c rapidsai/label/legacy-xgboost to the installation command, or add rapidsai/label/legacy-xgboost to the environment file’s channel list. Place this label before conda-forge when using strict channel priority. If the environment explicitly requests rapidsai::xgboost, replace that requirement with rapidsai/label/legacy-xgboost::xgboost.

The legacy label preserves existing packages; it will not receive new RAPIDS XGBoost builds. Keep XGBoost, RAPIDS, Python, and CUDA versions compatible when using these historical packages.