Requests enter/leave batch per token step. Eliminates idle GPU cycles from slow completions.

Tiles attention in SRAM. O(n) memory instead of O(n2). 2-4x faster on long sequences.

Non-contiguous physical blocks mapped to logical positions. 2-4x more concurrent sequences per memory budget.

40-50% VRAM reduction at near-identical quality. Enables larger batch sizes on same hardware.

Small draft model proposes tokens; large target verifies in parallel. Add last -- needs stable baseline.