[FEEDBACK] Reducing Docker Image Size by Decoupling Model Downloads

[addicted-ai]

Thank you for the great work and continuous development on this project. 🙌
I really appreciate the pace and quality of improvements.

Large Docker Image Size

I’ve noticed that the Docker images are steadily increasing in size as new features are added. This seems to be largely due to bundling all models directly into the image. While this makes the setup straightforward, it also means that after each updates where changes are primarily in the codebase, we still need to re-download a very large image, which could potentially be avoided.

Suggestion

• Check and download only the required models at container startup (this could be handled by the worker container; Immich follows a similar approach).
• Validate local model files by comparing size or a known hash to ensure completeness, and re-download only if validation fails.

[NeptuneHub]

Hi and thanks for the feedback. Last time I checked the image was around 6GB. I understand that is big and it is my mind the idea to reduce it.

The model themself are like 7-800mb and docker/Kubernetes already address the download of the container run layer: the only time it re-download everything is when big change are happening.

What I would like to check is:

• how the cached layer are configured in the Dockerfile, to try to avoid redownload as much as possible;
• look if I can reduce the footprint of the image themself: this is tricky due to the different dependency.

All of them are more a long time plan that I added in the tracker issue, this because change in this can have impact on the analysis itself (a different library version can produce different value in the inference) so need time to be tested.

Moving the download of the model in running time is not something that I really like. It it less than 20%, you still need to download and also add additional logic that can bring isse.

[crosbyh]

I am downloading the server container right now and it is pulling 11.05 GB and taking quite some time. Any efforts here would be much appreciated!