Ahmadzei's picture
added 3 more tables for large emb model
5fa1a76
The attention is computed from these separate vectors instead of a single vector containing the word and position embeddings.