Open
Description
The AutoSymbolics backend is internally calling modelingtoolkitze, which is very heavy handed and has a lot of bugs and limitations, e.g., around the type of the parameter object etc.. It would be nice if AutoSymbolics used Symbolics only to derive the gradient functions and let AutoModelingToolkit be a separate backend for those problems that are okay with being passed through modelingtoolkitze.
As an example of a bug this introduces, if p is a named tuple, modelingtoolkitze will somehow change that into an array, and that will not work in the functions that assumes p is a named tuple, giving an obscure error message.