Support OpenVINO vision model #32

This feature request aims to integrate support for vision language models into our existing framework. Currently, our framework supports non-vision models, but there is a need to extend this support to vision models, which are loaded and processed differently.

To achieve this, I have implemented the following:

1. Separated Initialization for Vision and Non-Vision Models:

   • Vision models are initialized using the ov_phi3_vision.py script provided by OpenVINO.
   • Non-vision models continue to be initialized using the existing methods.

2. Quantized Vision Model Support:

   • We have added support for the quantized vision model Phi-3.5-vision-instruct-int4-ov.
   • The model can be found at the following link: Phi-3.5-vision-instruct-int4-ov

Implementation Details

• Vision Model Initialization:

  • The OpenVinoEngine class now checks if the model is a vision model during initialization.
  • If the model is a vision model, it uses the ov_phi3_vision.py script to load and initialize the model.
  • The vision model is then processed using the AutoProcessor class from the transformers library.

• Non-Vision Model Initialization:

  • Non-vision models are initialized using the existing methods, ensuring backward compatibility.

• Streamlined Generation Process:

  • Both vision and non-vision models support streaming output, with timing information logged for performance analysis.
  • The generate_vision method has been updated to log prompt length, new tokens generated, time to first token, prompt tokens per second, and new tokens per second.

References

• OpenVINO Phi-3-vision notebook
• ov_phi3_vision.py