docs(ml): clarify cuda and openvino requirements #9072
Conversation
| - On Linux (except for WSL2), you also need to have [NVIDIA Container Runtime][nvcr] installed. | ||
|
|
||
| #### OpenVINO | ||
|
|
||
| - The server must have a discrete GPU, i.e. Iris Xe or Arc. Expect issues when attempting to use integrated graphics. |
There was a problem hiding this comment.
Expect issues when attempting to use integrated graphics.
What are the expected issues when using integrated graphics?
There was a problem hiding this comment.
Thanks for pointing me to those issues. I also notice myself that facial recognition started failing a few weeks ago when using OpenVino, when previously it worked just fine. I'm using Intel N100 which has integrated graphics. I had to disable OpenVino in order to get Immich working again.
What I don't understand from this new documentation is, do we expect these new issues to never be fixed, or is this just a temporary thing that we can't use OpenVino anymore?
If it is permanently broken, why does it no longer work? Did the model get too big or complex for the N100 integrated graphics? If so, what's your minimum specification?
There was a problem hiding this comment.
We updated OpenVINO and onnxruntime-openvino a month ago. That fixed two issues relating to external data and dynamic input shapes, but seems to have a major regression in iGPU compatibility.
As for fixing it, this is up to upstream in OpenVINO and/or ONNX Runtime. The prerequisites right now are just because of the current behavior in these libraries. I opened an issue about it and imagine the issue will be fixed at some point, but I don't know when.
Description
Specifies that OpenVINO should only be used with discrete GPUs since this is the current state of things. Also adds info about the minimum NVIDIA driver version.