Distributed XGBoost training on multi-node and multi-GPU cluster

Hi everyone,
I have set up a cluster of 2 nodes and each has 2 GPU through CLI using Dask, and I implemented the distributed xgboost training on this cluster, however, only one GPU utilized on each node. Does anyone know how to utilize both GPU on each node, totally 4 GPU, to train the xgboost model?

Could you please describe the way how you set up the cluster in more detail and share the code you have used?