[adaround][mix-precicison][onnx] how to manually set some nodes to fp16 when applying adaround on one onnx model?

[seriphdog]

Greetings.
I am dealing with a resnet_v2-7.onnx and I expect:

after applying adaround Alg. on all possible Conv/GeMM, this onnx will result in

1. all non-fused BatchNorm nodes set as fp16,
2. every Conv/GeMM set as W8A8/W4A8, (int4/8 here, not fp8 of course)
3. accuracy drop < 3%

Typically, I might utilize a MixPrecisionSimModel class , and reset dtype && bitwidth manually.

Is there an example to achieve above functions?

For aimet-onnx, QuantizationSimModel enables me manually set different int-width for given nodes, only if aimet_onnx/quantsim.py and other necessary files properly revised.

However, any suggestions for fp16 case? aimet-torch verifies a MixedPrecisionConfigurator class, not sure if same concept has already been implemented for aimet-onnx.