XGBoost now uses CTK-11.8 for binary packages, there's no need to maintain a cub submodule anymore.
XGBoost Plugin for Federated Learning
This folder contains the plugin for federated learning. Follow these steps to build and test it.
Install gRPC
Refer to the installation guide from the gRPC website.
Build the Plugin
# Under xgboost source tree.
mkdir build
cd build
cmake .. -GNinja \
-DPLUGIN_FEDERATED=ON \
-DUSE_CUDA=ON\
-DUSE_NCCL=ON
ninja
cd ../python-package
pip install -e . # or equivalently python setup.py develop
If CMake fails to locate gRPC, you may need to pass -DCMAKE_PREFIX_PATH=<grpc path> to CMake.
Test Federated XGBoost
# Under xgboost source tree.
cd tests/distributed
# This tests both CPU training (`hist`) and GPU training (`gpu_hist`).
./runtests-federated.sh