Quantization bert：How can I fix some layers for full accuracy?

[shiqingzhangCSU]

In paper “Understanding and Overcoming the Challenges of Efficient Transformer Quantization”，the authors point out that use the W8A32 or fixed layers full precision such as the residual connections can reduce accuracy loss. How can I use the quantization tool to fix some layers for FP32 quantization?

import onnx
from onnxruntime.quantization import quantize_dynamic, QuantType

model_fp32 = 'path/to/the/model.onnx'
model_quant = 'path/to/the/model.quant.onnx'
quantized_model = quantize_dynamic(model_fp32, model_quant)