is the docker image downloading all sorts of models?

[eleaner]

does it have to?
could the download be configurable?

[eleaner]

or maybe at least some ability to define different volume for models?

it is sort of possible currently but container won't start trying to take ownership of lost+found

[rishikanthc]

The container downloads all the nvidia models alone. The rest of the models are downloaded when you first use them.
For eg. If you select whisper small, the first run will download and cache whisper small. Subsequent runs will use the cached models.

The nvidia models alone (parakeet, canary and sortformer) are downloaded in advance.

It could be made configurable but that's a bit of work and not on my radar currently.

Right now, I've been working on performance optimizations, UX enhancements and a few new useful features and resolving certain core issues based on community requests.

This could be an enhancement I can work on once the popular things requested by the community have been incorporated.

However, more than happy to accept contributions from the community if someone has interest and time to implement this.

My time is limited so I tend to prioritize the more urgent and popular issues / features.

Edit: the main branch has changed which persist the downloaded models to an external volume to prevent redownloads on each restart / build. So that should help alleviate some pain. It will be out in the next release.

[eleaner]

you see
I do not even have nvidia
I am trying to run it on CPU the most effective way and many GB of nvidia models are kind of not that useful to me

I know the quality will not be that great but I am where I am

[eleaner]

I might end up having ROC so ... here is another idea of improvement :)