I tried updating to pytorch 1.12 and it did not fix the problem. It is possible to put the torch.multiprocessing.set_sharing_strategy("file_system") line just once in the torch replay buffer as that is always loaded in. However if the replay buffer is not needed in the future and that import is not longer always there other multiprocesses that use pytorch will likely break.
I tried updating to pytorch 1.12 and it did not fix the problem. It is possible to put the
torch.multiprocessing.set_sharing_strategy("file_system")line just once in the torch replay buffer as that is always loaded in. However if the replay buffer is not needed in the future and that import is not longer always there other multiprocesses that use pytorch will likely break.