Test a persistent GPU cache for slow first photo searches
At a glance
An Immich user traced first-search delays to a suspected CUDA compilation cache and tested persistence across container recreation. This workflow is listed under Self Hosting, AI Performance.
- Agents / tools
- Immich, Docker Compose, CUDA, nvidia-smi
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) In a test Immich installation with Docker Compose and remote CUDA machine learning, compare a first smart search with repeated searches while watching nvidia-smi. The reported setup used server and ML v2.7.5 with an 8 GB RTX 5050. 2) Inspect /root/.nv/ComputeCache inside the ML container while the first search runs; the user saw new folders during the slow search. 3) Add nvidia-test:/root/.nv/ComputeCache to that service's volumes and declare nvidia-test in the top-level volumes mapping. 4) Recreate the stack, perform the first search, then repeat it. Run docker compose down followed by docker compose up -d and repeat the comparison to check whether warm-search behavior survives recreation.
How it went
The user reports fast subsequent searches surviving recreation. A later driver update caused recompilation without observed cache-related errors.
Worth knowing
The first search remained slow. The author treated this as an experiment and did not recommend changing production solely on this evidence.
