immich: Hardware-Accelerated Machine Learning

[v3DJG6GL]

Is your feature request related to a problem?

Although immich is attached to my nvidia GPU, immich doesen't seem to use it for ML stuff: Rather than utilizing my GPU, my CPU is used for that stuff.

Describe the solution you'd like

There might have to be some yaml properties adjusted. See here: https://immich.app/docs/features/ml-hardware-acceleration

Describe alternatives you've considered

I can definitely live without GPU acceleration for ML. But since one GPU node is occupied for exactly that stuff, it would be kinda nice to have that GPU node also utilized :)

Additional context

I was asked to make this enhancement request on discord: https://discord.com/channels/830763548678291466/1221451548328984619/1221460825982636163

immich chart version: 14.2.4 OS: TrueNAS-SCALE-23.10.2

I've read and agree with the following

• [X] I've checked all open and closed issues and my request is not there.
• [X] I've checked all open and closed pull requests and my request is not there.

[stavros-k]

I don't think we have anything extra to do here. At least for NVIDIA, all it needs is the GPU added and capabilities set.

Both of those are done automatically already.

Note: The yaml files Immich has on their docs is about docker-compose configuration, not something immich it self needs.

[v3DJG6GL]

Here are the logfiles from immich-machinelearning pod: I am not sure but could it be that the wrong ml-image is used?

[03/30/24 23:28:36] INFO Starting gunicorn 21.2.0

[03/30/24 23:28:36] INFO Using worker: app.config.CustomUvicornWorker
[03/30/24 23:28:36] INFO Booting worker with pid: 78
[03/30/24 23:28:42] INFO Started server process [78]
[03/30/24 23:28:42] INFO Waiting for application startup.
[03/30/24 23:28:42] INFO Created in-memory cache with unloading after 300s
of inactivity.
[03/30/24 23:28:42] INFO Initialized request thread pool with 80 threads.
[03/30/24 23:28:42] INFO Application startup complete.
[03/31/24 13:32:19] INFO Setting 'nllb-clip-large-siglipv1' execution
providers to ['CPUExecutionProvider'], in
descending order of preference
[03/31/24 13:32:19] INFO Downloading clip model
'nllb-clip-large-siglipv1'. This may take a
while.
[03/31/24 13:32:19] INFO Setting 'antelopev2' execution providers to
['CPUExecutionProvider'], in descending order of
preference
Fetching 523 files: 100%|██████████| 523/523 [00:47<00:00, 11.06it/s] [03/31/24 13:33:08] INFO Loading clip model 'nllb-clip-large-siglipv1' to memory
[03/31/24 13:33:13] INFO Downloading facial recognition model 'antelopev2'. This may take a while.
Fetching 4 files: 100%|██████████| 4/4 [00:02<00:00, 1.35it/s] [03/31/24 13:33:16] INFO Loading facial recognition model 'antelopev2' to
memory
[03/31/24 13:37:23] INFO Shutting down due to inactivity.
[03/31/24 13:37:23] INFO Shutting down
[03/31/24 13:37:24] INFO Waiting for application shutdown.
[03/31/24 13:37:24] INFO Application shutdown complete.
[03/31/24 13:37:24] INFO Finished server process [78]
[03/31/24 13:37:24] INFO Worker exiting (pid: 78)
[03/31/24 13:37:24] INFO Booting worker with pid: 780
[03/31/24 13:37:29] INFO Started server process [780]
[03/31/24 13:37:29] INFO Waiting for application startup.
[03/31/24 13:37:29] INFO Created in-memory cache with unloading after 300s
of inactivity.
[03/31/24 13:37:29] INFO Initialized request thread pool with 80 threads.
[03/31/24 13:37:29] INFO Application startup complete.
[03/31/24 15:33:14] INFO Setting 'nllb-clip-large-siglipv1' execution
providers to ['CPUExecutionProvider'], in
descending order of preference
[03/31/24 15:33:14] INFO Loading clip model 'nllb-clip-large-siglip__v1' to memory

This is the immich-machine-learning that is used:

docker.io/altran1502/immich-machine-learning@sha256:612e8856ad92f5915ececd5a67b76b7218f0fe4e41c72fcf1675bee76802294b

And this seems to be the default CPU image...

[stavros-k]

Fixed in https://github.com/truecharts/charts/commit/0534cade2ac28c33a0fd3f1b3b097e48c80d9328

[truecharts-admin]

This issue is locked to prevent necro-posting on closed issues. Please create a new issue or contact staff on discord of the problem persists