I will have the model and data soon on github, but the issue here is not the overall slowness (takes hours). Perhaps if it took seconds I wouldn’t mind.
Rather, it is the fact that the first chain seems to manage better because the first 200 iterations are generated quickly to give the chain a head start, sort to speak. In the above chains, the rate for the first 200 iterations of the chains was about 1 it/s. This does not happen for the second chain - they take several orders of magnitude longer. Nor does it in any way help the model, because it seems clear that generating the “head start” iterations did help the model converge quicker.
But why shouldn’t it happen for the second chain as well?
I feel it is a bug, or at least something that should be configurable.