machine learning container using over 4gib ram
Unraid using docker compose community app. Is this normal I think it has changed after the new update?
20 Replies
Yes this is normal
The r requirement for Immich is 4-6
Machine learning will have to load the model into memory when it runs
Ok thanks
When does ML run, and are assets uploaded to the ML server? Sorry if this has been answered
ML run upon the asset is uploaded and jpeg thumbnail is generated
The thumbnail file is sent over http to the ml container
Got it thanks
Is the ML needed for extra features like your "map" etc ?
No it is not 🙂
Is it safe to limit the RAM usage to less than 4GB or would that cause instability?
it would crash the container becasue when the models are loaded it would exceed 4GB of RAM
Ok, and let's say the ML container is done processing some photos and sits idle for a while. Will the container free up that RAM to allow the host to reclaim that memory?
it will free up some of it
probably 60%
Ok and is the ML container required for Immich to run? Would it be possible to not use ML and still get all of the other features of Immich?
You'll need the ML container for all the features
you can disable ML by commenting it out. this would mainly affect search and facial recognition.
but in practice ML should use about 3-3.5gb under load. the higher usage is because of this bug https://github.com/immich-app/immich/pull/3207. try setting job concurrency to 1 for the ML jobs and restart your container.
GitHub
Build software better, together
GitHub is where people build software. More than 100 million people use GitHub to discover, fork, and contribute to over 330 million projects.
Just to confirm concurrency works as?:
1 give least resources to the job
5 give most resources
yes, it's the number of images processed at once for that job. higher concurrency is faster (to a point) but will use more resources
@sogan Ahh thanks for the link to the bug. Hopefully that gets fixed! Regarding concurrency - did you set the value of each of these settings in the UI to be "1"?
only
tag objects, recognize faces and encode clip need to be set to 1 to avoid the bug for now.Great, thanks, I'll try that!