NUTS and parallelized black box likelihood

Questions
2

I think that’s the only option, because the way pymc3 works (it ‘pickle’ the sampler and push it to multiprocess)
If you use cores=1, the logp and gradient is still using multiprocess tho.