The features X in our standard decision trees are float32, so it would make sense for the threshold of features to also be float32, see

Note that the Cython trees are expose in our trees, e.g. DecisionTreeRegressor attribute tree_.