Who's Looking at Whom?

In the decoder of an encoder-decoder transformer (e.g., the original machine translation model), the cross-attention sublayer computes attention where the queries come from the decoder's own sequence, but the keys and values come from where?

The raw token embeddings of the target sequence before any layers

The encoder's output representations of the source sequence

The decoder's own previous layer output (same as queries)

A separate learned embedding table unrelated to either sequence