Model trained on GPU always loads data to GPU when predict

Hello, I am using the XGBoost sklearn API, specifically the DaskXGBRegressor, to train xgb models using multiple GPUs on a single machine. What I noticed is that when I used the regressor to predict, it always loaded some data on to the first GPU core. I extracted the Booster obj from the regressor and use that to predict. It still loaded data to the first GPU. I checked the “predictor” attribute of the regressor and it is None. Is there any way I can force the regressor or the booster to predict only using CPU? Thank you very much.

BTW, I saw somewhere that if I saved the Booster to disk then loaded it back, then it could predict using only CPU. I’d rather not taking the round trip is possible. Thanks again.

XGBoost version: 1.7.3