Pickling

Python’s pickle module relies on being able to import classes using the fully-qualified name: module_name.class_name. This means that any class must be defined in the module’s global scope to be pickled successfully.

Classes dynamically generated (e.g., by FitConfig(...).fit_class or FitConfig.from_...) are not automatically registered in the global scope, which makes them unpicklable by default.

To fix this, you have two options:

1. Register the class in the global namespace manually:

   import sys
   setattr(sys.modules[__name__], 'MyClientClass', generated_class)
   

2. Or create a named subclass in the global scope:

   class MyClient(FitConfig(...).fit_class):
       pass
   

Either approach ensures the class has a stable, importable name that pickle can resolve when deserializing.

Or you can use an automated tools like dill or cloudpickle, which are more forgiving in these scenarios.