I think torch.distributed.init_process_group() can be executed since there is no error even in the world_size=1 state.

By allowing this to be executed even in the world_size=1 state, it is possible to check the operation with respect to the function assuming that torch.distributed.is_initialized() is True, without having to run MPI.