ReadonlyinputThe input embeddings [1, T, hidden] (the first segment boundary). The caller owns and disposes it.
Release the pass's caches and masks.
Layers [lo, hi) over h, returning the residual stream after layer
hi-1 (no final norm) and the fetched K/V of the reused donors in the range.
h is never disposed; donorKvIn holds earlier segments' donor K/V
(caller-owned). The caller owns and disposes what comes back.
One sequence streamed through the layer stack in segments. Borrows the ids it began with; owns its caches, masks and per-layer inputs until
dispose.