ONNX Runtime doesn't support the graph optimization of vision-encoder-decoder yet

[CrasCris]

Is there another way to quantized TrOCR model to speed the inference, i alreayd have in ONNX form but the inference time is to high