Hi, I am using a TRUENAS-MINI-3.0-X+ running the latest scale 26.0 beta3 (64GB RAM, 1 apps pool, 1 data pool).
I was looking into immich and the possibility to use machine learning to tag images. I understand that a dedicated GPU (the more VRAM, the faster the better) would help - but it is feasable without the GPU (the trueNas Mini does not come with one) and it “only” takes longer. Or is it not possible at all?
According to the documentation I can setup a dedicated immich machine learning docker container on a different computer and use this - could this be done on a Macbook Pro with a M1 or do I have to abandon the idea without a newish GPU altogether (yeah, the current hardware prices are not exactly encouraging)
Any opinions or experiences are helpful. Thanks in advance
Not an opinion, but evidence-based: It’s perfectly fine. I’m running Immich on the test host in the signature within a Bhyve Linux VM. (Alongside some other ‘microservice’ VMs.) The CPU is an eight-core Atom (Denverton class). It just takes a bit more time. But no general restrictions.
for the m1 remote idea: run the immich machine-learning container on the macbook and point the mini Immich at that host (remote machine learning in the docs). m1 wants the default/cpu image there, not cuda. if the library isnt huge overnight on the mini cpu is usually enough and u can skip the remote box.
I think I will give it a try with overnight then. Are there any recommendations in regard to the settings? Which model to use? Immich settings under machine learning settings? What would be the URL - default the value seems to be “http://machine-learning:32002” - should I change that to the IP of my immich (and is the port correct - the GUI is at 30041?)?
leave http://machine-learning:32002 alone if Immich and the ML container are on the same TrueNAS app. thats the docker hostname on the app network, not the web UI.
30041 is only the Immich GUI port. dont point ML at that.
for overnight on the mini cpu id keep the stock models (CLIP + face) and maybe drop ML concurrency a notch if the box gets warm. only change the URL if u move the ML container to another machine.
