Re: the memory aspect you can control what variables pymc actually stores in the trace with var_names so that’s not a problem if they’re just nuisance parameters. You do loose the rhat and convergence statistics ofc.
Re: iterative cdfs, we only have a few for like hypergeometric, and incomplete beta and gamma, which are directly on log-scale (IIRC). I think we copied them mostly from stan-math, except for one where we found an alternative algorithm that was slightly more stable. So we’re behind you guys on that front.
As mentioned above we don’t even have the mvnormal cdf.