Per-layer decoder state: self-attention K/V [B, T, D] (un-split, un-scaled, exactly what the oracle concatenates) and the cross-attention K/V computed once from the encoder output.
Per-layer decoder state: self-attention K/V [B, T, D] (un-split, un-scaled, exactly what the oracle concatenates) and the cross-attention K/V computed once from the encoder output.