Llama 3.2 Vision: 32 self-attention layers read only text-token KV; 8 cross-attention layers read only image-token KV. PagedAttention reserves (T+I)×40 slots; the ideal is 32T+8I.
-slots reserved
-slots needed
-wasted fraction (computed)
self-attn × text (read)cross-attn × image (read)reserved, never read
Hover a cell.
Two-level allocator
image-type small pagestext-type small pagesToy: 256 B and 384 B → 768 B large pages
Per-layer prefix hits Illustrative
Click KV tokens or Mamba checkpoints to evict/restore them. A model-wide hit needs every layer type satisfied at the same prefix. (The paper checkpoints every 512 tokens; here every 4.)
Waste fraction vs image-token share, from the episode's slot formula (reserved 40(T+I), needed 32T+8I). Computed, not measured
Measured on this model: 1.71× throughput. At 0% image tokens the formula still gives 20% (cross-attention layers reserve text slots); the episode notes the chat-mix sweep was never run.