State is not copied for parallel chains
When sampling multiple chains in parallel, the Markov chain is copied to the executing process. I assume the initial state is thereby reset to the starting point given by the problem or the starting_point
argument.