This message was deleted BentoML #ask-for-help

Join Slack

This message was deleted.

# ask-for-help

Slackbot

12/05/2022, 4:51 PM

This message was deleted.

Aaron Pham

12/05/2022, 8:10 PM

can you send your bentofile.yaml here?

Aaron Pham

12/05/2022, 8:10 PM

it would help if you can post the stack trace here as well.

Greg

12/06/2022, 2:17 AM

Copy code

service: "service.py:svc"
include:
  - "service.py"
  - "configuration.yaml"
docker:
  distro: debian
  base_image: "<http://gcr.io/my-ai-org/debian-py39-cuda116-conda:latest|gcr.io/my-ai-org/debian-py39-cuda116-conda:latest>"
  setup_script: "./setup.sh"
  env:
    BENTOML_CONFIG: "src/configuration.yaml"

Greg

12/06/2022, 2:18 AM

No available stack trace, aside from imports failing from packages installed in the

base_image

Greg

12/06/2022, 2:19 AM

I can install via

setup.sh

, but on every

bentoml containerize

, you have to rebuild those dependencies, which is incredibly slow for some CUDA-enabled libraries.

Aaron Pham

12/06/2022, 3:12 AM

Hmm this might has to do with buildkit caching. I assume you just run with bentoml containerize?

Aaron Pham

12/07/2022, 2:55 AM

right now I believe that conda is not cache correctly with docker. Essentially we are creating a conda environment everytime we are running

containerize

. This means that if the base image already have a conda environment, the result container from bentoml containerize would not have access to the conda env from the

base_image

Aaron Pham

12/07/2022, 2:55 AM

one improvement we can make is that for base image that has conda, we can use that conda env instead of creating a new one? cc @Sean

3 Views

Open in Slack

Previous Next