There is a work-around to access the underlying XGB Booster:
booster = model.get_booster() dtest = xgb.DMatrix(X_test) y_shap = booster.predict(dtest, pred_contribs=True)