core/nn/transformer_decoder library
Stack of TransformerDecoderBlocks with an optional final
LayerNorm.
This is the "seq2seq decoder" (masked self-attention + cross-
attention + FFN, per block). It takes a decoder-side hidden state
x and an encoder-side memory and returns contextualized
decoder representations of the same shape as x.
No token/positional embeddings and no LM head — the caller is
expected to prepend those. See EncoderDecoderTransformer for a
full seq2seq wrapper.