Dask for Training but not for Predictions

My Training dataset is big enough to require dask, but my test dataset fits on one GPU no problem.

To speed up the predictions, is it possible to use just regular GPU predictions without using dask?

The answer seems to be yes, the training outputs a dict so you need to call the ‘booster’ from the training dict and then it functions as a normal xgboost model.

Maybe this wasn’t supposed to work but it seems to…