To continue our discussion - We’d be interested in finishing the deployment of (higher-order) AD and testing framework, at least to the point where it wouldn’t interfere with something like including RHMC into Stan.
Regarding mixed-mode testing - @bbbales2, @Bob_Carpenter, so, if I understand correctly, AD testing is under review, but does not does not test mixed-mode/higher-order? But you you think it would be relatively easy to extend what has been done to higher-order testing?.. And in terms of fwd support, the basic functions/pdf’s are supported, but most of the advanced stuff isn’t (solvers, integrators, etc…)?
As @betanalpha pointed out there are also two things that need to be dealt with in Stan itself:
- how to gracefully fail when you use something without forward mode support and
- how to compile higher-order support only when needed, because it would be slow.
Any ideas/comments on these would be appreciated. Is there something else that needs to be done?