I'm using xgboost for quantile regression, and for that, I have to construct a xgboost.QuantileDMatrix object for training.
Based on the documentation, it seems that I need to keep the original QuantileDMatrix used for fitting the models as a reference (the ref parameter) when making predictions. This is extremely impractical, because my prediction environment is much lighter than the fitting environment and the memory limits become a concern with the heavy fitting data.
Is there any workaround or other solutions for this?