Clarification regarding Quantization in ExecuTorch

ExecuTorch
1

Hi ExecuTorch Team,

image
image1309×901 138 KB

In ExecuTorch stack as part of Program Preparation, we have optional quantization step.
If want to plug in our own quantization mechanism, is there any customization entry points for that?
Please guide if possible, how we can achieve this?

2

Hi @rohit-kumar-manav

Yes, you can write a custom quantizer based on your capability of your backend.

Here are some references:

https://pytorch.org/executorch/stable/quantization-overview.html
https://pytorch.org/tutorials/prototype/pt2e_quantizer.html

1 Like