silu(gate(x))·up(x) for exactly one activation row over matching axis-1
geometry (L at most 12, at most 8192 bf16 inputs; the bit widths may differ). tail applies where the
widths differ; same widths take the float32 sigmoid. Borrows x and the
weights; returns owned lazy output with x's leading dimensions.
silu(gate(x))·up(x) for exactly one activation row over matching axis-1 geometry (L at most 12, at most 8192 bf16 inputs; the bit widths may differ).
tailapplies where the widths differ; same widths take the float32 sigmoid. Borrows x and the weights; returns owned lazy output with x's leading dimensions.