I have a model trained that uses a lightgbm classifier with a sklearn calibration on top.
I am tracking the entire training process via mlflow as well as registering the final model to then be loaded in a deployment.

from lightgbm import LGBMClassifier as lgbmc
from sklearn.calibration import CalibratedClassifierCV

mlflow.start_run()
mlflow.<flavor>.autolog()

model = CalibratedClassifierCV(lgbmc(learning_rate=args.learning_rate, n_estimators=args.n_estimators, subsample=sr, class_weight='balanced', num_leaves=args.num_leaves), cv=args.cv_folds, n_jobs=-1).fit(x_train, y_train)

mlflow.<flavor>.save_model(
        lgb_model=model,
        path=args.model
    )
mlflow.<flavor>.log_model(
        lgb_model=model,
        registered_model_name=args.registered_model_name,
        artifact_path=args.registered_model_name,
    )
mlflow.end_run()

Both flavors seem to work in this case but which flavor should be used in such multi-flavor ensembles or stacked models?