use single point prompt mask decoder slow than pytorch, how to solve this problem?

[stihuangyuan]

[AlessandroMondin00]

I have runned the same experiment too: with a Macbook Air M2 it takes the following times to extract image embedding with ViT-L (sam_vit_h_4b8939):

1. ~36 seconds with pt weights
2. ~42 seconds with onnx encoder extracted with samexporter.export_encoder
3. ~ 39 seconds with onnx encoder quantized extracted with samexporter.export_encoder