heads. Those are the attention weights from every token with global attention to every token
            in the sequence.
    Nr2