That’s a good option too. Sometimes the model consists of some training and inference code, which can be loaded as a library, plus a bunch of weights, which may

That’s a good option too. Sometimes the model consists of some training and inference code, which can be loaded as a library, plus a bunch of weights, which may be large enough to warrant a separate file. Either way, I think the deployment of the model isn’t really a hard problem. Validation of model quality, keeping track of what code was used, and what training data, making sure the updated data is where it’s supposed to be before training, tying all these parts together in some kind of comprehensive management system are all harder problems I think.