A module turns a complex architecture into a repeatable unit. In a Transformer, the unit is a block that alternates an attention sublayer with a two-layer feedforward network, then repeats through the model. section 9.1 section 15.6
A Transformer stack
Picture a sequence passing through several identical-looking blocks: attention selectively pulls information from elsewhere in the sequence, and the feedforward sublayer transforms the resulting representation. Stacking these modules builds a deeper model without inventing a new architecture at every depth. section 15.6 section 9.1
Use these distinctions when reading an architecture diagram:
- Treat a module as a reusable architectural pattern, not merely as one isolated layer. section 9.1
- Expect Transformer modules to pass information through attention and then a two-layer feedforward network. section 15.6
- Read the stack as a hierarchy: repeated modules organize the model while attention provides sequence-wide information flow. section 9.1 section 15.6
