Transformer Architecture Refresher

A map of the decoder-only transformer behind Llama 3.1 8B: the embedding, 32 identical blocks, attention, the gated MLP and position, with every shape.