KFold example with `ddp_spawn` crash

[niberger]

🐛 Bug

When using the cross validation loop from the example pl_examples/loop_examples/kfold.py with ddp_spawn strategy, it encounter a SIGABRT Exception and the program crash.

To Reproduce

import os
from pytorch_lightning import seed_everything, Trainer

from pl_examples.loop_examples.kfold import KFoldLoop, LitImageClassifier, MNISTKFoldDataModule


def run():
    seed_everything(42)
    model = LitImageClassifier()
    datamodule = MNISTKFoldDataModule()

    trainer = Trainer(
        default_root_dir=os.getcwd(),
        limit_train_batches=1,
        limit_val_batches=1,
        limit_test_batches=1,
        num_sanity_val_steps=0,
        max_epochs=1,
        enable_model_summary=False,
        strategy="ddp_spawn",
    )
    internal_fit_loop = trainer.fit_loop
    trainer.fit_loop = KFoldLoop(5, export_path="./")
    trainer.fit_loop.connect(internal_fit_loop)
    trainer.fit(model, datamodule=datamodule)


if __name__ == "__main__":
    run()

Expected behavior

Training without exception

Environment

• CUDA:
  • GPU:
    • Tesla P100-PCIE-16GB
    • Tesla P100-PCIE-16GB
    • Tesla P100-PCIE-16GB
    • Tesla P100-PCIE-16GB
    • Tesla P100-PCIE-16GB
    • Tesla P100-PCIE-16GB
    • Tesla P100-PCIE-16GB
    • Tesla P100-PCIE-16GB
  • available: True
  • version: 10.2
• Packages:
  • numpy: 1.20.2
  • pyTorch_debug: False
  • pyTorch_version: 1.9.0
  • pytorch-lightning: 1.6.0rc0
  • tqdm: 4.60.0
• System:
  • OS: Linux
  • architecture:
    • 64bit
  • processor:
  • python: 3.8.9
  • version: Proposal for help #1 SMP Fri Jan 14 13:59:45 UTC 2022

Additional context

I am working on a fix.

cc @Borda @carmocca @justusschock @ananthsub @ninginthecloud @rohitgr7