BentoML

Hi <@U011UKLMP6Z>, yes you are right, it is linked to Bento during bentoml containerize now

You can perhaps try mounting the `~/bentoml/models` directory to the build container

so that the build process can fetch models from the model store, where you have the pytorch models saved

alternatively installing BentoML on the build container and run `bentoml containerize` should work as well, not sure if that’s hard to do with your current setup

This is a optimization step we did for 1.1. If you need to you can always export the bento first, then containerize it that way

Your first suggestion worked, <@UKB4CLKP1>, thank you! For the second suggestion, it would be quite hard, because the container used is basically one with "docker in docker". It doesn't even have Python installed. I wonder if it would be possible to have an option for `bentoml containerize` to make it "dry-run". Instead of running it, it would create return the temp dir and the docker command. It would make it easier to integrate into CI/CD, which commonly gives you a special container with Docker

Oh, does the export put everything into a dir?

wdym by “temp dir and the docker command”?

From what I read from the code, it seems that it creates a dir, puts everything there and then invokes the docker build

export will copy the model into the artefact

yes `bentoml.container.build` generate the final dockerfile, and builds everything from the tempdir

The dockerfile includes within the bento is only a minimal dockerfile to make it compatible with non buildkit setup

Hmmm. I'm currently deploying the model with that dockerfile. Do you think I'll have a problem with it?

the reason for tempdir is for buildkit support, which is opt-in

If you want to get the final dockerfile that includes all of the feature bentoml containerize support, you can try `bentoml.container.get_containerfile`

this will return the dockerfile we will use for the final `containerize` step