It looks like there is a planned upcoming deprecation of the ML framework APIs in BentoML. Just curious if there is any GitHub issue or documentation about this that we could look at as a preview of what is to come. Seems similar to the deprecations from 1.2.0 but then brought back in a later 1.2.x release, so I’d be interested in getting a sense of the rationale and what the upcoming alternative will be. Thanks! 😃

Hi Adam, we don’t have a docs for this yet, but the plan going forward is that for most frameworks users should use

bentoml.models.create

API to save to BentoML store

with bentoml.models.create("my-model") as model:
  # your saving logic here

In users training script, etc. Then in your service definition, you can use the following pattern

@bentoml.service
class Service:
  model = bentoml.models.BentoModel("my-model")

  def __init__(self): ...

If you are using HuggingFace models then you can use

bentoml.models.HuggingFaceModel

directly

@bentoml.service
class Service:
  model = bentoml.models.HuggingFaceModel("my-model")

  def __init__(self): ...

You can see more here https://docs.bentoml.com/en/latest/guides/model-loading-and-management.html#model-loading-and-management