Not sure on the cause but if you do:
mp.set_start_method('fork', force=True)
There is no semaphore leakage
now it has log, but only the initial one
So the subprocesses are not logged ?
Let me check, see what can be learned ...
Let me know if it solved it, if it did I'll make sure we push the RC
and one experiment takes 40 hours to run, so i let them run in parallel
Wouldn't it make sense to use a single one ?