[Feat]: Minified UI/configurator-only install to use OneTrainer outside of main compute nodes?

[rafa-br34]

Describe your use-case.

So I have a very specific use case for OneTrainer where the actual training happens in a headless Linux environment. The issue is that the whole OneTrainer install (as of 07254ad) needs 166+ Python packages, taking up way too much space (4.44GB on Windows) and time to install.

What would you like to see as a solution?

The most obvious solution would be a WebUI (easy to use and move around), but the clear consensus is that a WebUI version isn't going to happen. So the closest option is to use the UI locally and export configs (or use X11 forwarding if you're insane).
Now, I haven't checked the codebase myself much, but I'd guess it uses an architecture similar to Kohya's (a configurator + the actual training scripts).
If that's the case, I think this project would benefit greatly from a "configurator only" install that doesn't need 4+ GB of storage and 166 packages.

Have you considered alternatives? List them here.

No response

[dxqb]

what you are asking is 3 things:

• remotely controlling a CLI OneTrainer. That is already implemented, and imho in a better way than you are proposing. look at the cloud tab
• a minimal install. @hameerabbasi I think this would be possible with the options like [cuda]?
• a web UI might happen, there is an unreleased one by @BitcrushedHeart. There is also this PR that changes the existing UI to Qt6 implement Qt6 views #1446. The UI isn't what's causing 166 dependencies though.

[hameerabbasi]

a minimal install. @hameerabbasi I think this would be possible with the options like [cuda]?

Yeah; it's possible, but it also needs changes on the code side. We'd need to separate the UI imports from the torch/transformers/diffusers etc. imports. We'd also need to move the training imports into the training functions; and/or guard with a try ... Except ImportError: ... if missing.

[hameerabbasi]

One could also make a helper utility that does the try/except thing.

[rafa-br34]

what you are asking is 3 things:

* remotely controlling a CLI OneTrainer. That is already implemented, and imho in a better way than you are proposing. look at the `cloud` tab

* a minimal install. [@hameerabbasi](https://github.com/hameerabbasi) I think this would be possible with the options like [cuda]?

* a web UI might happen, there is an unreleased one by [@BitcrushedHeart](https://github.com/BitcrushedHeart). There is also this PR that changes the existing UI to Qt6 [implement Qt6 views #1446](https://github.com/Nerogar/OneTrainer/pull/1446). The UI isn't what's causing 166 dependencies though.

I only found out about the cloud tab later, but that's a very clean solution.
I saw the Qt6 issue, and I agree Qt6 would be better. (Although I personally didn't have any issues with the current UI so far)