I expect to start the process with torch spawn method and import boxx

[yanziwei]

I minimally reproduced the startup process:

import torch
from boxx import *

mp = torch.multiprocessing.get_start_method(
            allow_none=True)
print(mp)

I expect to get None or spawn The actual got the fork

Background:

self.sampler = DataLoader.Sampler()
mp_start_method = torch.multiprocessing.get_start_method(
            allow_none=True)
if mp_start_method is None:
            torch.multiprocessing.set_start_method('spawn')

I expect CUDA tensors to be shared between different processes.

[DIYer22]

神奇的 bug, 我毫无头绪 🤣