That error makes sense because for this model the marginal likelihood is actually a lower bound to the true marginal likelihood (if you’re using the default VFE approximation). Instead of using MvNormal it’s calculated using pm.Potential. That means you’ll have to calculate it from scratch.
I’d have to double check, but I think the DTC and FITC approximations have a proper likelihood, not a lower bound, that they use. That might help with the calculation.