Once your model parameters are decided upon, it is time to train the final model, the one that will get deployed in production. I will shortly explain how this final model will be different than the ones you train while developing your model.

Data splits

Dataloader and Dataset
You do not split into train/val/test datasets anymore, you train on the entirety of the data.

Early stopping

Early Stopping (patience)
Early stopping itself is a hyperparameter. You just train to the amount of epochs that was deemed optimal during early stopping.

Cross Validation

K-Fold cross validation tunes hyperparameters over all folds, therefore you get only one set of hyperparameters which you use on the final model. No issue at all.

Nested cross validation

When using Nested cross validation, you would have n different sets of hyperparameters. You would then train n models (on the entire dataset) and your final model would be an ensemble method over these n models.