Hitting a weird error to do with RNGs in Scan in a custom function inside a Potential

[ricardoV94]

@jessegrabowski

It uses an index sequence instead of scanning over the variables themselves, but my (admittedly uninformed) intuition is that these should generate equivalent graphs in the end.

Unfortunately, those are completely different, because the grad can’t see the graph outside of scan when defined like that. That’s stated explicitly in the docs (second box):

https://pytensor.readthedocs.io/en/latest/tutorial/gradients.html#computing-the-jacobian

I fixed that in my WIP refactor of Scan because it’s extremely unintuitive.

Anyway there is a known bug when you try to get a gradient of a Scan wrt to something that used to have RVs but no explicit updates. I never managed to narrow down that but here is where it was last discussed:

Potential bug using scan to create random variable in PyMC

Anyway I’ll try to actual check it locally