train in ipynb with accelerator="auto" failed

[Borda]

🐛 Bug

Is expected to run auto mode of PL Trainer in ipython notebook on a machine that has two GPUs?

To Reproduce

Run a Trainer with

accelerator="auto"
devices="auto"

In a Jupyter notebook cell.

MisconfigurationException: `Trainer(strategy='ddp_spawn')` or `Trainer(accelerator='ddp_spawn')` is not compatible with an interactive environment. Run your code as a script, or choose one of the compatible strategies: Trainer(strategy=None|dp|tpu_spawn). In case you are spawning processes yourself, make sure to include the Trainer creation inside the worker function.

Expected behavior

just take the best supported and run, no crashing

Environment

• PyTorch Lightning Version (e.g., 1.5.0):
• PyTorch Version (e.g., 1.10):
• Python version (e.g., 3.9):
• OS (e.g., Linux):
• CUDA/cuDNN version:
• GPU models and configuration:
• How you installed PyTorch (conda, pip, source):
• If compiling from source, the output of torch.__config__.show():
• Any other relevant information:

Additional context

cc @justusschock @kaushikb11 @awaelchli @ninginthecloud @akihironitta @rohitgr7

[awaelchli]

For accelerator="auto", devices="auto" we didn't consider that in Jupyter notebooks, we need a different default for the strategy. So in this sense, strategy should also be selected automatically. For multi-gpu in notebooks, only strategy="dp" works here. So if we want to support that, we would need to add a condition to the AcceleratorConnector.

[nicocheh]

Hi @awaelchli , i used DDPSpawnStrategy(find_unused_parameters=False) with accelerator="gpu" and some weeks ago it worked in a jupyter notebook using multiple GPUs. Did something changed that is not working any more?

[awaelchli]

This is fixed now with #13405. accelerator="auto", devices="auto" will select the right strategy in the jupyter notebook.
@nicocheh For setting find_unused_parameters_false, we added the option strategy="ddp_notebook_find_unused_parameters_false".