Sampling doesn't start when njobs > 1 for some models

The multicore sampling is completely rewritten by now, so if there still is an issue with frozen chains, it clearly is different. Are you using the most recent pymc3 release? Does this happen sporadically or regularly with one model (if so, which)?