The cache a composition's KV scheme names for one sliding-window layer of
maxSize (cache index layer): plain for bf16, for a layer a per-layer
scheme does not list and for turbo (TurboQuant encodes full-attention
layers only), the bf16-first rotating affine lego otherwise.
The cache a composition's KV scheme names for one sliding-window layer of
maxSize(cache indexlayer): plain forbf16, for a layer a per-layer scheme does not list and forturbo(TurboQuant encodes full-attention layers only), the bf16-first rotating affine lego otherwise.