What runs where?

[ZdsAlpha]

I am so confused at this point what runs where! I am trying to install dependencies using setup_sh but those dependencies are not found when code runs. I cannot use requirements.txt since I have to manually install and uninstall some dependencies. Also does it create conda environment and I am installing dependencies outside of it.
All I am trying to do is install mmdetection and run my custom model artifact. Wasted entire week trying to solve dependencies issue. Documentation is useless, doesn't go in detail or a little depth when not using mainstream repos.

[parano]

@ZdsAlpha what's the reason you need to manually uninstall?

For advanced use cases, we typically recommend using your own custom docker base image, that way you can prepare the docker image any way you want: https://docs.bentoml.org/en/latest/concepts.html#custom-docker-base-image

It is also possible to use the conda environment in user-provided setup_sh, as long as you are installing to the default base environment, e.g. conda install -y -n base YOUR_CONDA_PACKAGE

Could you share your service definition code and setup_sh code so we can help look into it?

[ZdsAlpha]

Yes the content of file are 3.7.11. I am using bentoml/model-server:latest-py37-gpu image. why does it have python 3.8 installed in it?

[ZdsAlpha]

I have mounted the image bentoml/model-server:latest-py37-gpu and checked python version and it is 3.8.8. Isn't this a bug?

[parano]

That's indeed a bug in our docker release script https://github.com/bentoml/BentoML/blob/v0.13.1/docker/model-server/release.sh#L68 This only affects base images for GPU.

The issue was not discovered because, for most users, bentoml's docker build script will try to re-install the correct python version, although it breaks your use case when the setup_sh script is also installing to conda env.

I'm working on a hotfix now to re-publish all GPU-enabled docker base images for 0.13.1.

[ZdsAlpha]

Thank you very much!

[parano]

@ZdsAlpha The images are published to docker hub now, I've verified the python version issue has been fixed. Could you try again and let me know?