This is a classical problem that I would guess almost every PyMC3 user faces at some point or another… the simplest and most common workaround is to wrap your val in a pm.Normal with a very small sd parameter value (but not so small that you get the “Bad inital energy: -inf” error message).
If you are more strict about the exactness of the model you infer, I’ve read that it may be possible to solve this problem with a custom likelihood using pm.Potential – but I don’t have any experience with that (yet).
A thread of potential interest (since the participants talk about this problem a bit): Help with Multivariate SDE Timeseries