It needs a mlx fork because the lowest bit in mlx is 2 currently (for affine quantization).