OCR Performance question

[AI091]

Hi team, I see that in Docling v2.56, the default OCR model is changed to Rapid-OCR. What are the implications of this with GPU acceleration, is it well supported? I tried it earlier but it seemed onnxruntime didn't use the GPU properly.

[dolfim-ibm]

If you install onnxruntime-gpu it should work.

[AI091]

Thanks, trying it out

[AI091]

I am still getting very high CPU usage , I think its because of this RapidAI/RapidOCR#112

[AI091]

should I be using torch backend instead?

[AI091]

Running with openvino works properly, even slightly faster than EasyOCR with GPU accleration.

[dolfim-ibm]

Thanks for giving it a try. We will make sure to document it. (if you would like to make a docs contribution, you are welcome)

[AI091]

Thanks, will try to do it soon :)

[dolfim-ibm]

@AI091 which steps did you take to run the openvino version? Some extra dependency?

[AI091]

Yes, just installed openvino package and chose it as the backend in RapidOcrOptions