The choices qualified on architecture: the single-row representations on applegpu_g13s only, wide prefill on GPU family 15 (M3) and newer. A layer resolves this once, when it is constructed, and passes the stored result to the kernels on every call.
architecture
applegpu_g13s
The choices qualified on
architecture: the single-row representations onapplegpu_g13sonly, wide prefill on GPU family 15 (M3) and newer. A layer resolves this once, when it is constructed, and passes the stored result to the kernels on every call.