feat(machine-learning): support cuda 12

[martabal]

Add support for cuda 12:

• Onnxruntime supports it since 1.17 (docs)
• The default cuda version with ubuntu 23.10 is cuda 12.
• Keep the default version to cuda 11

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Pages

Latest commit:	1a8ddde
Status:	✅  Deploy successful!
Preview URL:	https://05e22c0c.immich.pages.dev
Branch Preview URL:	https://feat-support-cuda12.immich.pages.dev

View logs

[mertalev]

I'm not sure we actually need to keep 11. In practice, 525 is the minimum driver version for ONNX Runtime based on issues I've seen and the minimum compute capability for it is 5.2. I think anyone who's using CUDA right now is already in an environment where 12 will just work.

[martabal]

the original goal was only to support cuda 12 but if you want to move away from cuda 11 completely, I'm all in 😄

[martabal]

What do you think about still supporting cuda 11 ?
It could be useful non-docker users so they still can install it with something like poetry install ... --with cuda11

[mertalev]

I'm not sure if Poetry can handle different flavors of the same package since the optional dependencies are all squashed into the same lock file. It didn't work for cpu/gpu variants of PyTorch at least. Only one of them actually get resolved.

[mertalev]

But that might have changed since I tried it. You can test and check if it can actually differentiate the cuda-11 and cuda-12 flavors

[martabal]

Yeah, you're right, we can't python-poetry/poetry#7748