Immich ML Hardware Acceleration with OpenVINO in a Proxmox LXC container

Also known as: I am stupid.

What?

(This is hopefully going to be a short one)

I recently installed Immich as I’m trying to get away from backing up my pictures on someone else’s computer, or on NextCloud (awesome tool, horrible auto-sync behavior in the Android app, no fancy features for photo hosting).

I chose to run it in an LXC container because it allows me to let it use my GPU without having to do PCIe Passthrough, which reserves the card for that machine¹. I’m also running the default Docker Compose setup (yeah, yeah, container in a container… I don’t care).

The official instructions from the Immich team are on point: I uncommented all the things, chose the correct backend, did all the things… But still, every time I checked, the only available ML provider was CPUExecutionProvider.

Why?

I actually delayed starting using Immich until I figured this out, but after many reinstalls, in multiple separate occasions, installing the intel-opencl-icd in all the affected machines², and still never seeing the OpenVINOExecutionProvider even being available I had all but given up.

So I did what any sane person would do, and removed all the configs³ and started from scratch. And that’s when I (finally?) spotted this comment in Immich’s docker-compose.yml:

  immich-machine-learning:
    container_name: immich_machine_learning
    # For hardware acceleration, add one of -[armnn, cuda, rocm, openvino, rknn] to the image tag.
    # Example tag: ${IMMICH_VERSION:-release}-cuda
    image: ghcr.io/immich-app/immich-machine-learning:${IMMICH_VERSION:-release}

What?

For hardware acceleration, add one of -[…] to the image tag

I tried it, it worked instantly⁴.

At which point, after writing a whole paragraph about “why is this not in the official instructions?” I checked them for what feels like the 75th time⁵, and saw it:

3. Still in immich-machine-learning, add one of -[armnn, cuda, rocm, openvino, rknn] to the image section’s tag at the end of the line.

While I don’t agree with this being step 3, when step 2 involves uncommenting and changing lines that are below it in the docker-compose.yml file, or the fact that they could name the default image cpu, noaccel, or something that makes it more obvious that it needs to be changed, I am definitely nitpicking here.

The underlying point, which I also made at the very start of this, remains: I am stupid. But hey, at least my Immich ML works now.