Save downloaded models in Docker

argosopentech · December 31, 2022, 7:37pm

github.com/LibreTranslate/LibreTranslate

Ability to provide already downloaded language models instead of slowly downloading them at first run

opened 08:11PM - 27 Nov 22 UTC

closed 04:57PM - 20 Dec 22 UTC

It would be nice if I can specify a directory with an --argument where already t…he downloaded files are in. I noticed that the initial download at the first run is very slow (often below 1 Mbit/s) and it needs to do it on every installation again. To reduce the load and traffic on the already slow servers it would be nice if just one download would be necessary instead of one on every deploy.

Do we have an example Dockerfile for saving downloaded models? We could put this in the default Docker example since most people will want to avoid redownloading on every Docker run.

pierotofy · December 31, 2022, 7:54pm

In the README we actually already recommend people use run.sh (https://github.com/LibreTranslate/LibreTranslate/blob/main/run.sh) (and run.bat on Windows) which take care of saving the models to a permanent storage location when running with docker.

Ablaze1528 · January 2, 2023, 11:58am

check this: Only download specific language models? - #4 by argosopentech

according to that info, you could mount a volume to have the downloads persist?

    volumes:                                                                                                                                        
      - ./data/argos-translate:/home/libretranslate/.local/share/argos-translate

I’m just not sure how frequently those change, i.e. does it make sense to persist the models?

argosopentech · January 2, 2023, 3:05pm

The models change pretty infrequently so in most cases it’s probably fine to just download and leave them. I’m currently working on functionality to upgrade only the models that have newer versions.