Bounded Replay

An optimization mechanism used by inference engines when processing sliding-window attention layers. KV cache is retained only for the limited history inside the window, avoiding wasted GPU memory on tokens outside the window and reducing redundant computation during the prefill stage.