The purpose of validation data is to find the best hyperparameter combination that makes your model generalize to unseen data. You can assume that your optimal hyperparameter combination does not depend on the specific combination of training & validation sets.

So, you would use the same number of epochs while training the final model.