Setting gradient attributes to Trainer when using manual optimization

[clement-pages]

Hey lightning team,

I am currently working on adding manual optimization to pyannote.audio in this pull request. The goal is to let the user choose between manual and automatic optimization by setting model.automatic_optimization. In our code, we want to retrieve gradient attributes (for instance gradient_clip_val or gradient_clip_algorithm) from the trainer by doing model.trainer.gradient_clip_val for example.

The problem is that we cannot instanciate a Trainer object with gradient_clip_val if we set automatic_optimization to False. Lightning throws the following exception:

lightning_fabric.utilities.exceptions.MisconfigurationException: Automatic gradient clipping is not supported for manual optimization. Remove `Trainer(gradient_clip_val=5.0)` or switch to automatic optimization.

So my question is : Is there any way to instanciate a trainer with gradient attributes and manual optimization without rising this error ? Maybe it could be a warning message instead of an error ? I have found a work around, but not having this error would clearly simplify the code.

Have a nice day!