[Feature] Facial recognition concurrency

[Zhanyuanium]

I have searched the existing feature requests, both open and closed, to make sure this is not a duplicate request.

• Yes

The feature

TL;DR: Facial recognition is too slow without concurrency.

Current facial recognition concurrency value is limited to 1, leads to poor performance and waste of powerful hardware.

I deployed a machine learning server with Nvidia RTX 4070 Ti, for there are hundreds of thousands of photos to process. The face detection concurrency was set to 3 to fully utilize the VRAM, led to rapid procession less than 4 hours, produced millions of faces. However the facial recognition concurrency could only be set to 1, caused low GPU utilization and very slow procession, which is estimated to last a few days. If this could be concurrently processed, huge performance improvement would be expected.

Sorry I am a noob. I think maybe there is a global write lock in the table related to clustering, which disables concurrency.

Any ideas, insights or advice please? Thanks!!!

Platform

• Server
• Web
• Mobile

[bo0tzz]

The algorithm is single threaded, this fundamentally has to run 1 at a time.

[mertalev]

Try restarting the container. It checks if reindexing should be done at startup. If this is a new immich setup and you imported that many assets, then the indexing right now will be very suboptimal until that restart.