A finer-grained part inside a module, such as an attention head or projection layer.
What It Is
Components include query/key/value projections, individual attention heads, layer-normalization cells, and gating units inside a larger block. They are the smallest units this taxonomy names before dropping to raw tensor ops.
Why It Matters
Clear component language helps you read architecture diagrams, follow sharding plans, and understand which part of a module changes when papers propose a new projection or norm placement.
Simple Example
Multi-head attention is a module; each head is a component with its own Q/K/V projections and output merge. Grouped-query attention changes how those components share key/value state.
Common Confusions
Components are not standalone modules—you rarely swap only a norm and call it a new architecture. They are also not parameters: a component is a structural role, while weights are the learned values filling that role.