Workaround to avoid unexpected stop of execution
Unfortunatelly, there are some limitations related with using together: multiprocessing and multithreading (forking, handled by omp). As a workaround, graph is not shared, each worker has own instance. There is also possibility to change multiprocessing mode to 'spawn' (and this case was checked) and it works with omp multithreading, but it results in same bahavior: model is copied in memory (because OS cannot postpone copying according copy-on-write strategy - which is available in 'fork' mode).