No error crash, just a never ending pause

Nobody has claimed this yet.

### Bug description

Notebook - https://www.kaggle.com/code/vigneshwar472/baseline-residualunetse3d
Github repo (ResidualUNetSE3D implementation) - https://github.com/wolny/pytorch-3dunet/tree/master

## Issue - Training starts on 1x P100 GPU but it does not start on 2x T4 GPU
I want to use 2 GPUs simultaneoulsy for training (ddp_notebook strategy) But I do not know, training does not start and 2 GPUs were not in use

I have no idea "why it's not working".

Check the error messages and logs section.

### What version are you seeing the problem on?

v2.4

### How to reproduce the bug

```python
Go the Kaggle Notebook https://www.kaggle.com/code/vigneshwar472/baseline-residualunetse3d-train

Copy & Edit
Run All
You will encounter a never ending pause
```

### Error messages and logs

### When used with 2x T4 GPUs
```
n = len(folds)

for i in range(n):
print(f'fold {i} started....')
model = ResidualUNetSE3D(in_channels=1, out_channels=6)
lm = CZIILightningModule(model=model)
logger = CSVLogger(save_dir='/kaggle/working/training_results', name=f'fold_{i}')
trainer = Trainer(accelerator='gpu',
strategy='ddp_notebook',
devices=2,
precision='32',
gradient_clip_val=None,
logger=logger,
max_epochs=15,
enable_checkpointing=True,
enable_progress_bar=True,
enable_model_summary=False,
inference_mode=True,
default_root_dir='/kaggle/working/training_results',
num_sanity_val_steps=0)
trainer.fit(model=lm,
train_dataloaders=DataLoader(folds[i][0], batch_size=1, num_workers=4, shuffle=True),
val_dataloaders=DataLoader(folds[i][1], batch_size=1, num_workers=4, shuffle=False))
del model, lm, logger, trainer
print(f'fold {i} completed....')
```


### When used with 1x P100 GPU

[
](url)

### Environment

Please go the kaggle notebook and run it

### More info

_No response_

cc @justusschock @lantiga

Start by running the linked Kaggle notebook with v2.4 and compare the 1x P100 and 2x T4 configurations. Inspect the Trainer setup using strategy='ddp_notebook' and devices=2, along with the provided logs; no local file or test is named. Done means reproducing the pause, identifying its cause, and confirming a working multi-GPU run.

Written by the indexing model from the issue text.