optimum-intel
optimum-intel copied to clipboard
huggingface
Pipeline
Hi @echarlaix . I just recognized that we didn't have any ipex optimization guide in README.
To make it more friendly to users, I integrated the generation modeling in our new pipeline. Users can just import pipeline from optimum-intel and easily apply optimum-intel optimization by only changing 1 line code from HF.
I think this approach is better than the old one, and it also aligns with optimum-nvidia
You can see that I simplified the pipeline as much as possible because we only need to consider text-generation for now. With this framework, it would be better to enable other tasks in the future.
Would you please help to review these changes? Thx!
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
Hi @echarlaix I have updated the pipeline design to try to align with the optimum. Please take a look and check if the design is correct. Once our design is done, we could discuss the tests : )
Hi @IlyasMoutawwakil . Thanks for your review. Would you please check my new changes? Thx!
@jiqing-feng great PR! I left a comment about one last place where hub kwargs should be passed as well.
Can you please add a test for the default task models, to make sure they're always usuable/exportable with simple pipeline(task=)
Hi @IlyasMoutawwakil @echarlaix . I have finished your required changes, but the CI fails are not related to my changes. Would you please take a look? Thx!
Hi @IlyasMoutawwakil , I have added the test for the default task models, but these models are too large, so the CI is over time. Do you have any idea how to fix it?
Thank you @jiqing-feng , yeah I didn't expect that, let's not include them for now. We can add them to a slow tests CI later. otherwise LGTM !
Hi @IlyasMoutawwakil . I think we could merge this PR since the failed CIs are not related to my changes
To fix the code style test you can do the following :
pip install .[quality]
make style
Hi @echarlaix . I have fixed all your comments except the model/graph check. Would you please review it? Thx!
Hi @echarlaix @IlyasMoutawwakil . Thanks for your review. I suppose we can merge it.