fix: DRY + speed up docker build

[akx]

Description

This PR:

• Refactors the Dockerfile to use an in-container setup script instead of commands being mashed into RUN statements.
• DRYs out the repeated package lists, PyTorch package index URL determination etc.
• Switches to using uv for installing Torch as well (it's faster!)
• Adds a check that the functionality to pre-load whisper and embedding models actually works

This should have no effect for the end user other than a possibly slightly smaller image. For developers, this is easier to maintainer. For CI and builders, this is faster.

Testing & review

I checked that an image built with and without USE_OLLAMA works as before.

I didn't check the CUDA configuration, since I have no CUDA-enabled Docker hardware at hand right now.

---

Changelog Entry

Changed

• Optimized Docker build a bit.

[Yanyutin753]

Hey, bro, I think it's in the dev branch ，not the main branch

[justinh-rahb]

@akx Love to see some work on making Docker builds faster but I will note that we've had some past issues getting uv to work reliably for all build platform combinations, and for PyTorch at all. Please ensure you've thoroughly tested this 🙏

[tjbck]

Thanks for the PR, LGTM but more testing wanted here!

[akx]

Hey, bro, I think it's in the dev branch ，not the main branch

I couldn't find any instructions on which branch to target, but I noticed a lot of work happening on both dev and main.

[akx]

@justinh-rahb:

we've had some past issues getting uv to work reliably for all build platform combinations

What are the Docker platform combinations to target here? linux/amd64, linux/arm64? Others?

I can write e.g. a Makefile to build all images. :)

, and for PyTorch at all.

uv is moving pretty fast right now (disclaimer: I'm a (very) minor contributor) and I see its version hasn't (and isn't, even with this PR) pinned, so chances are it now does work – very much seems to be able to be imported, anyway:

$ docker build -t owu-builtin-ollama . --build-arg="USE_OLLAMA=true" && docker run -it owu-builtin-ollama python -c "import torch; print(torch.__version__)"
 => => writing image sha256:56268c22bee4e9edfc507cd7358945e4acd7f4f32e1c358dcb0f03067acdc1a7                                                                                                                                                  0.0s
 => => naming to docker.io/library/owu-builtin-ollama                                                                                                                                                                                         0.0s
2.3.0

(same results repeated for no builtin ollama)

Is there a particular e.g. CUDAness scenario in which it hasn't worked here? Is there something I could manually test that'd certifiably use Torch within the image?

@tjbck

LGTM but more testing wanted here!

As above: anything you'd particularly like me to test (within the constraints here, i.e. I'm working on Apple Silicon, no CUDA here)?

[justinh-rahb]

I'll go through our convos the last time I messed around with uv installing torch for which platform I had issues with.

[tjbck]

We would need testing with all 6 of our variants!

[justinh-rahb]

And test runs of Github action workflows

[justinh-rahb]

@akx I think there's been some signifcant changes as of late to our build workflow. If this PR is still applicable I encourage you to keep working on it, or if we've resolved some of the original issues that prompted you to create it perhaps it can now be closed.

[akx]

@justinh-rahb The repetition that this PR was fixing is still there in the Dockerfile as far as I can see. I'll maybe revisit this once someone takes a look at my other PRs (#2041, #2233) – it's discouraging to keep rebasing them to no review or interest.

[justinh-rahb]

The repetition that this PR was fixing is still there in the Dockerfile as far as I can see. I'll maybe revisit this once someone takes a look at my other PRs (#2041, #2233) – it's discouraging to keep rebasing them to no review or interest.

I understand your frustration but please also remember that we're also volunteers that have regular jobs too, and the project has been moving very quickly. I am personally trying to get the backlog looked at right now, thus why I've been checking in with the open PRs and communicating with our contributors in the backchannels to get things moving along.

[akx]

I understand your frustration but please also remember that we're also volunteers that have regular jobs too, and the project has been moving very quickly.

Absolutely, I'm in the same position. Sorry if I sounded a bit unkind there :)