Thanks @jiqing-feng but the quicktour looks a bit crowded already, I don't think we want to add more code snippet for now : would make more sense to have a description explaning when to use ipex vs openvino vs neural-compressor (+ a link towards a benchmark could be very nice also) as it's currently a bit confusing for users what should be used when + what are the integrations for each. I think it also makes sense to keep all code snippet related to optimum-intel in the dedicated section to not duplicate them multiple times (same applied for others sections) we can just add links in the quickstour instead

Agree, do you think we should use optimum-intel pipeline instead? In this case, users can pass accelerator=ipex;ov;neural-compressor to choose which kind of backend they want to use. The descriptions should be:

To apply optimum-intel optimization, you just need to import pipeline from optimum-intel instead of transformers, and choose accelerator from ov(link), ipex(link), neural-compressor(link) to get different acceleration. Following codes:

- from transformers import pipeline
+ from optimum.intel.pipelines import pipeline

model_id = "distilbert-base-uncased-finetuned-sst-2-english"

  # Run inference!
- classifier = pipeline("text-classification", model=model_id)
+ classifier = pipeline("text-classification", model=model_id, accelerator=accelerator)
results = classifier("He's a dreadful magician.")

If we introduce optimum-intel in optimum quicktour in this way, it could be more clean and easy to understand for users.
Do you mind update the readme like this? If so, please open a new PR cause I don't have the right to delete OV context. Thanks!

Thanks @jiqing-feng but the quicktour looks a bit crowded already, I don't think we want to add more code snippet for now : would make more sense to have a description explaning when to use ipex vs openvino vs neural-compressor (+ a link towards a benchmark could be very nice also) as it's currently a bit confusing for users what should be used when + what are the integrations for each. I think it also makes sense to keep all code snippet related to optimum-intel in the dedicated section to not duplicate them multiple times (same applied for others sections) we can just add links in the quickstour instead

@echarlaix , we're fine w/ the "add link" idea.
For positioning, do you think it helps to position ipex + inc as intel's HF HW-accelerated PyTorch-native runtime, no need for proprietary format/conversion/... ? For inc, we position it as intel's quantization lib(like autoawq etc. in ecosystem), but have intel enhancement algorithms like autoround, and we are planning to deprecate its runtime role and move runtime role back to ipex backend. So, IPEXModel can run both models quantized by ecosystem libraries(autoawq, gptqmodel etc.) and models quantized by inc.
It needs time, but it's what we will do in 2025.