@tqchen, I think we should handle the weights more generally. Here I was referring to binary artifacts that are produced (which are not present in the relay graph initially as weights do) as part of the lowering of the external function that is required in the runtime.