Decode integer codes back to a float32 approximation of the original.

dequantize is the inverse of quantize — the read side of the affine map. It undoes the offset and scaling to recover a floating-point tensor from the stored integer codes. It is the operation that lets a quantized weight participate in an ordinary float matmul: the sidecar-design quantized modules keep their weight as int8 codes, then call dequantize on every forward to reconstruct a float weight before the GEMM. The result is not the exact original — it is the original rounded to the nearest grid point — but the error per element is bounded by s / 2.

The decode is the affine inverse:

where the subtraction removes the zero_point offset z and the multiply restores the physical step s. As in quantize, a per-channel scale / zero_point (given ch_axis) is reshaped to broadcast along the channel axis, so each channel is decoded with its own step. The whole computation is a subtract-then-multiply on the public lucid.* op surface (H4), identical on the Accelerate and MLX streams.