Instead, the sequence is typically broken into subsequences equal to the model's maximum input size.