Full Attention

The standard attention mode of the Transformer decoder, in which each decoding step computes attention weights over all historical or reference tokens; information is lossless, but computational cost grows with sequence length.