My own embedding about tensorflow-seq2seq-tutorials HOT 1 CLOSED

(I assume you are talking about output projection layer, i.e. softmax layer)
embedding_size is the dimensionality of word vectors, not amount of different words decoder can produce. In the end you need to output token IDs, not embeddings, so you have to get from decoder's hidden state to probability distribution over possible words, i.e. vocabulary. That is, softmax layer.

I imagine you could try decoding embeddings directly, then searching for a word via cosine similarity vs whole vocabilary, but this would be waaaaay more computationally expensive then softmax. Softmax is already the most expensive operation in this architecture.

from tensorflow-seq2seq-tutorials.