And should the Potential add a lot of overhead / compute time for NUTS? It’s so slow. I’m on PyMC 5.3 and the MvNormal usage you see above (with Poisson likelihood) is all I have – do you think updating to the latest PyMC 5.10 will speed things up?
Have you tried an alternative NUTS backend (blackjax, numpyro, or nutpie)?
No, the Potential should add little to no overhead