🐛 Bug

The following program never terminates.

import torch
import torch.multiprocessing as mp


def foo():
    x = torch.ones((2, 50, 10))
    return torch.einsum('ijl,ikl->ijk', x, x)


if __name__ == '__main__':
    foo()
    p = mp.Process(target=foo)
    p.start()
    p.join()

The behavior persists if one changes the einsum inside foo with an equivalent operation (e.g., bmm(y, y.transpose(1,2)), or (y.unsqueeze(2) * y.unsqueeze(1)).sum(3).

It doesn't reproduce, however, if one doesn't call foo inside the main block.

Environment

PyTorch version: 1.0.1.post2
Is debug build: No
CUDA used to build PyTorch: 9.0.176

OS: Arch Linux
GCC version: (GCC) 8.2.1 20181127
CMake version: version 3.13.4

Python version: 3.7
Is CUDA available: No
CUDA runtime version: No CUDA
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA

Additional context

Perhaps related to #2245.