I don’t have a specific model for this, but it’s happened to me a decent amount and folks report it here pretty frequently too:
I’ll have a model where all the parameters are on scales where weakly informed priors assign mass mostly within the -3:3 range and constraints that should yield finite likelihood for any possible combination of parameter values, yet default initialization fails for all chains. Can anyone explain why this can happen? Possibly something nuanced about the gradient that wouldn’t be obvious?
If it helps at all, at least for the cases I’ve personally encountered, most of the time I can get initialization to succeed by specifying either init=0 or init=x, with x being some value smaller than the default value 2.