Questions about the ‘optimize.torch.quantification’ qat mechanism

[Feynman1999]

❓Question

According to my understanding, coremltools.optimize.torch.quantification encapsulates the official quantification of torch (as an extension of FX Graph Mode Quantization in PyTorch).
My question is, must I use the coremltools API for qat training? Can I use the PyTorch's API directly to complete qat and convert it to an mlmodel?
In other words, what specific operations were done by coremltools.optimize.torch when completing qat for coreml package?
（I am still a newbie in quantification. Thank you for your guidance）

[aseemw]

Yes, you can use PyTorch's quantization APIs directly and convert to Core ML. That will also work.
However, that may not be the most performance optimal. The default settings used by torch's quantizer (in terms of symmetric asymmetric quantization modes and which ops are quantized) are not optimal for Core ML stack and Apple hardware (see details in this section). If you use coremltools.optimize.torch the correct default settings will be applied automatically.