But the model ultimately still has to process the comma, the newline, the "job". Is the main time savings that this can be done in parallel (on a GPU), whereas in typical generation it would be sequential?