ReadonlycodesReadonlydeltaReadonlygeometryReadonlyrepresentationsReadonlyscalesReadonlyspecReadonlyuseReadonlyvariantThe decode variant every kernel over these weights runs (kernels/trellis/index.ts): the row-form layers built on this projection take it at their construction.
A caller may prove row-contiguous, aligned input from an allocating op
such as RMSNorm. Lazy array strides cannot establish that before eval.
Rows past the tiled kernels expand the dense weight and multiply; that
product is evaluated before returning, so building the next projection
never holds a second expansion, unless boundExpansion is false: a caller
that queues several projections at once (the Neural Engine split's GPU
share) holds their expansions until it evaluates them.
StaticisStaticloadOptionalrepresentations: TrellisRepresentations
What the kernels take from this machine's GPU family (single-row representations, wide prefill), fixed here and passed on every call.