They almost certainly only do greedy sampling. Beam search would be a lot more expensive; also I'm personally skeptical about using a complicated search algorithm for inference when the model was trained for a simple one, but maybe it's fine?