← back

block-evict block-aligned eviction

16×8 paged-KV grid (128 blocks). Each block holds 16 token slots + a mean future-attention score. Click a block to toggle low/high attention. Eviction scores whole blocks and frees the lowest — never splits a block, never breaks COW pairs.

Click block = toggle attention 0.05 ↔ 0.95 · score = meanAttn − 0.1·frag

Freed blocks

0 / 128

Fragmentation

–

COW pairs retained

0

No COW pairs yet — press Fork to clone a block copy-on-write.

high attention (keep) low attention (evictable) freed COW-shared (pinned) evicting…

Block scores (lowest first, top 16)

blockmean attnoccupancyfragscorestate

score = meanAttn − 0.1 × frag, frag = 1 − occupancy/16. COW blocks are pinned (score = ∞) so the fork child survives eviction.