Python multiprocessing.Process hangs when large PyTorch tensors are initialised in both processes

1

Why does the code shown below either finish normally or hang depending on which lines are commented/uncommented, as described in the table below?

Summary of table: if I initialise sufficiently large tensors in both processes without using "spawn", the program hangs. I can fix it by making either tensor smaller, or by using "spawn".

Note:

  1. All memory is purely CPU, I don’t even have CUDA installed on this computer
  2. This issue does not occur if I replace torch with numpy, even if I make the array size 10x larger
  3. Version information: Ubuntu 22.04.1 LTS, Python 3.10.12, torch 2.1.2+cpu
Uncommented Commented Behaviour
(1), (4) (2), (3), (5) Hang
(2), (4) (1), (3), (5) OK
(1), (5) (2), (3), (4) OK
(1), (3), (4) (2), (5) OK 
import multiprocessing as mp
import torch

def train():
    print("start of train")
    x = torch.arange(100000)            # (1)
    x = torch.arange(10000)             # (2)
    print("end of train")

if __name__ == "__main__":
    mp.set_start_method('spawn')        # (3)
    x = torch.arange(100000)            # (4)
    x = torch.arange(10000)             # (5)
    p = mp.Process(target=train)
    p.start()
    p.join()