When using mlflow to train and register a text classifier model using the package, we often have to write our own pyfunc wrapper, which could be tedious.
The idea would be to provide a generic wrapper (with some hypothesis on the stored artifact) usable by all package users.
As a bonus, it could be really great if this wrapper was self sufficient (no need to install the package for inference, only torch and numpy)