C4A Attention Pattern

Hover over query tokens or compressed entries to see the connections

Pattern Description:
• Each token produces: q, k(v), Ca & Za, Cb & Zb
• q → k(v): Window attention (128 tokens, shown as all tokens here)
• C^compress is produced from Ca & Za, Cb & Zb
• q → C^compress: top-512 selection of compressed entries within range
Query (q)
Key-Value k(v)
Ca & Za
Cb & Zb
C^compress
Sink Token