A KV scheme as a table entry serves it: its storage, the kind and widths. A
composition's scheme also names the affine read kernels (kernels) and the
delayed conversion start (quantizedKvStart); the cache composes both from
the composition, so neither constrains the graph.
A KV scheme as a table entry serves it: its storage, the kind and widths. A composition's scheme also names the affine read kernels (
kernels) and the delayed conversion start (quantizedKvStart); the cache composes both from the composition, so neither constrains the graph.