ReadonlygpuReadonlyheadReadonlylinearReadonlyprogramReadonlyrestReadonlyspecReleases this layer's program and frees its weight rows; a shared program or buffers outlive it while another layer holds them.
StaticbuildLoads aneAffineSplitProgram(K, N, fraction, rows, slices) for pieces of
up to rows rows now, as AneTrellisMlpSplit.build loads its program.
share: as AneTrellisMlpSplit.build. gpuRows: the layer over the GPU's
rows.
Optionalshare: AneAffineSplitOptionalslices: numberOptionalgpuRows: AneGpuRows
Affine projection split: the first n output rows of an MLX affine projection run on the ANE from fp16 rows the GPU dequantizes straight into the streamed program's weight buffer; the GPU computes the remaining rows; outputs concatenate in row order.