Yup that’s right. @wmbelk are you having the problem described in this issue (that I forgot about…)?
For HSGP you can use a X = pm.MutableData, but as described in the gp.prior_linearized docstring. For gp.prior and other GP implementations, I guess you maybe “could” mutate the input data (I can’t think of a reason to), but if you’re doing it to get predictions from the GP it won’t give you the right answer.