Local AI backends for OPUS-MT and other HuggingFace models.

[Nigel2392]

I'm wondering; is anyone interested / would it be possible in this project to integrate local LLM backends? (Of course I will publish my code with tests)
These might be useful for developers wanting to translate locally; without cost. (There are some really low resource models out there!)
Allows for more leniency - but the drawback, most translation models I could find which are reasonably performant do not support HTML translation.

I currently in my own project have the following block which only allows for the local AI backend if transformers is installed - wondering if this would be the recommended way of implementing it.

try:
   from transformers import (...)

   class OpusTranslator(...):
      ...
except ImportError:
   OpusTranslator = None

[chris48s]

wagtail-localize has support for several machine translators already:

• deepl
• google cloud translate
• libre translate

each of these is an implementation of BaseMachineTranslator

All 3 of these implement the translate() method by sending text off to an external API and returning the result, but there's no reason why the translate() method couldn't do the translation using a locally installed LLM.

I'd suggest having a look over https://github.com/wagtail/wagtail-localize/tree/bb04c7627701c7922be66389a1aaa2a721a09559/wagtail_localize/machine_translators

Sounds like what you want to do could be achieved by making an implementation of BaseMachineTranslator, which can then be used in WAGTAILLOCALIZE_MACHINE_TRANSLATOR["CLASS"].

[Nigel2392]

Yes! I do already have this - the OpusTranslator is already a subclass of BaseMachineTranslator. Just wondering if anyone would be interested. :)

[zerolab]

Build it and they will come ;)
I am quite sure if you package it and share in the Wagtail Slack, folks using wagtail-localize will at the very least play with it

[zerolab]

To add to what @chris48s shared, I'd suggest making that its own module/package as you can iterate faster when it is on its own.