Putting already trained model artifacts into production

[ptaban]

Hey, so if i have model artifacts(lets say pickle files) that are already trained and ready to be served, should i juse mlflow.pyfunc to create wrapper which reads those artifacts and implement predict function, which i register to mlflow registry, so i dont take care of that artifact anymore, its gone.
Or another solution is to write class into production code that reads the artifact and does inference as inherited from mlflow api to customize it.

So first option is for example take notebook, upload pickle files, create wrapper, initialize wrapper, register to mlflow registry. And i dont care about those pickle files, since when reproducing model artifacts will now come from mlflow run.

Second option is to put pickle files in safe place, and read it on runtime of inference pipeline.