It’s not the same issue as what we’ve seen in Mac and windows. If it were, @mishooax should be seeing broken pipe errors instead of not seeing the sampling begin.
Linux still defaults to forking new processes with the multiprocessing module, whereas the issue in the other OS’es happens because they spawn new processes instead.
The most likely reason that the sampling doesn’t start has to do with memory usage. When sampling is performed on multiple cores, the entire model and its observations are copied to every new process. With models that are slightly big and have a large number of observations, the process of copying may take a long time or may even crash.
1 Like