Does PyTorch 2.1 Support Learnable Post-Training Quantization?

quantization
1

I’m exploring PyTorch 2.1’s quantization features and have a question: Is learnable post-training quantization (PTQ) supported in this version? Specifically, I’m interested in adjusting quantization parameters, like scales, while keeping other model weights fixed.

Thank you for your assistance and insights.

2

yes, it’s possible to support this, you’ll need to write your own quantizer though: How to Write a Quantizer for PyTorch 2 Export Quantization — PyTorch Tutorials 2.1.1+cu121 documentation and you can use this to specify the LearnableFakeQuant: https://github.com/pytorch/pytorch/blob/main/torch/ao/quantization/quantizer/quantizer.py#L59
https://github.com/pytorch/pytorch/blob/main/torch/ao/quantization/_learnable_fake_quantize.py

I’m not sure what learnable PTQ means actually

1 Like