Welcome!
It is not entirely clear what is going on here. If you have a runnable example (e.g., something simpler than what you are currently using) someone might be able to investigate further.
A couple of suggestions. First, I would install the most recent version of PyMC if you can mange to do so. Second, I would suggest avoiding explicit specification of the step methods and allow pm.sample() to automatically select step methods (including compound methods). There is not particular reason to use Metropolis except in relatively specific scenarios. If NUTS works for you model, you should probably use it. Third, 80,000 samples is likely excessive unless your model is sampling very very poorly (at which point you should just fix your model).