Immich — community fork (Blackwell / CUDA 12.8)

Caution

⛔ Deprecated — do not use this fork

Use the official image instead: ghcr.io/immich-app/immich-machine-learning:release-cuda.

This fork and the ghcr.io/ekropotin/immich-machine-learning:cuda-blackwell image are unmaintained and should not be used. The premise below turned out to be wrong: the CUDA 12.8 rebase was never necessary. The official image already works on NVIDIA Blackwell (sm_120) — onnxruntime ships PTX that the NVIDIA driver JIT-compiles to sm_120 at runtime (verified on an RTX 5060 Ti, back to onnxruntime 1.19.2). The CUDA failure 35 this fork tried to fix is a GPU driver / container passthrough problem on the host, not a problem with the image. Fix the host GPU passthrough and the stock image works.

Full write-up: immich-app/immich#28583.

Everything below is kept for historical context only.

This is a community fork of immich-app/immich that exists for one purpose: to provide an immich-machine-learning image that works on NVIDIA Blackwell GPUs (RTX 50-series, compute capability sm_120) until upstream support lands.

Why this fork exists

The upstream release-cuda ML image is built on nvidia/cuda:12.2.2-runtime-ubuntu22.04. CUDA 12.2 predates sm_120, so on Blackwell hardware ONNX Runtime silently falls back to CPU with no error — the only visible symptom is 0% GPU utilization during smart-search / face-detection jobs.

This fork bumps the prod-cuda stage to nvidia/cuda:12.8.1-runtime-ubuntu22.04 and works around a packaging quirk where libcudnn9-cuda-12 drags in cuda-cudart-12-2 and overwrites the CUDA 12.8 runtime that the base image provides.

Tracking issue upstream: immich-app/immich#28031.

See the full diff vs upstream: immich-app/immich:main ↔ ekropotin/immich:fix/cuda-blackwell-support.

Scope

• Only the machine-learning/ service is modified.
• Server, web, mobile, and everything else are kept in lock-step with upstream via periodic rebase.
• Once upstream ships an ML image with CUDA 12.8+, this fork will be archived.

Using the published image

# docker-compose hwaccel override
services:
  immich_machine_learning:
    image: ghcr.io/ekropotin/immich-machine-learning:cuda-blackwell

See machine-learning/README.md for build and usage details.

Upstream

For everything else — features, installation, docs, demo, contributing — refer to the upstream project:

• Repository: https://github.com/immich-app/immich
• Documentation: https://immich.app
• Discord: https://discord.immich.app