[Bug]: INT8_PTQ & INT8_ACQ Performance and Accuracy for WinCLIP

[junxnone]

Describe the bug

INT8 Performance

As the NNCF document says, when quantizing a Transformer model, the parameter model_type should be set to TRANSFORMER.
In my tests, if it is not set, the performance of the INT8 model will decline. If it is set, there will be a 10% performance improvement.

• PTQ Quantize
• ACQ Quantize

INT8 Accuracy

• Image-level seems to have a small reduction
• The accuracy at Pixel-Level has decreased so much that it cannot be used.
• No matter whether the parameter preset is set to MIXED or PERFORMANCE, it is the same.

Dataset

MVTec

Model

Other (please specify in the field below)

Steps to reproduce the behavior

• Export WinCLIP OpenVINO INT8 Model(PTQ & ACQ)
• run benchmark with the OpenVINO model

OS information

OS information:

• OS: [e.g. Ubuntu 22.04]
• Python version: [e.g. 3.10.14]
• Anomalib version: [e.g.1.2.0.dev0]
• PyTorch version: [e.g. 2.4.0+cpu]

Expected behavior

works for the INT8 Model

Screenshots

No response

Pip/GitHub

GitHub

What version/branch did you use?

No response

Configuration YAML

None

Logs

None

Code of Conduct

• I agree to follow this project's Code of Conduct