This message was deleted.

The work around would probably to not use

bentoml.model

. So when you download from S3, instead of using the bentoml model, you would just use the normal frameworks (i.e. torch, tensorflow)

Will it still be able to leverage the runner features? And does dumping model to bentoml repo provide any advantage over directly using?

yes, you will still be able to use the runner as usual. This a design choice of BentoML and we plan to remove the model dependencies from the build step in the future

Great, also I had model details [model_name] in bentofile.yaml and I think this resulted in the requirement of model to be earlier while running the command bentoml build. When I removed that from bentofile.yaml, it worked and the service was running. Since it was running with bentoml serve service.py, is this correct explanation that since in serve since we aren't giving bentofile.yaml it doesn't require the model to be there and can download it on runtime but if we provide in the bentofile.yaml it checks whether the model is there or not? [Edit : Not sure, while running container it is giving error]

yes

Okay great. Last question, currently it is giving me permission errors to save the bento model as "bentoml" user doesn't have access to the /home/bentoml/models/. Do we have any options to override either user or change the directory of model saving via bentofile.yaml?