Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

additional gains from QDQ #11260

Closed
mengniwang95 opened this issue Apr 19, 2022 · 6 comments
Closed

additional gains from QDQ #11260

mengniwang95 opened this issue Apr 19, 2022 · 6 comments

Comments

@mengniwang95
Copy link
Contributor

Hi,

I am confused about what additional gains we can get from QDQ format compared with quantization with QLinearOps, can you share me some ideas? For example, if QDQ format is more general, which cases confirm it?

Thanks

@yufenglee
Copy link
Member

Comparing with QLinearOps, QDQ format is much more flexible and helps the ONNX quantization ecosystem. Here are some benefits examples:

  • different EPs/accelerators supports different set of quantized operators. As long as they support QuantizeLinear and DeQuantizeLinear, they can run the model successfully.
  • ONNX exporter (like pytorch to onnx) and converter (like tf2onnx) can convert/export quantized models from original framework to quantized ONNX model with QDQ format.

@BIT-jwq
Copy link

BIT-jwq commented Aug 30, 2023

Hello Yufeng,

@yufenglee I am still curious to learn more about this topic.

  1. Is the advantage of this format solely related to supporting different EPs? Are there any other benefits associated with this format?

  2. Where can the theory paper or official demonstrations be related to the QDQ format?

I would appreciate your comments on this matter.

In fact, I am planning to conduct comparative experiments to validate the advantages of the QDQ format.

Could you please provide some advice and share your comments on this?

@Rickustc
Copy link

Hello Yufeng,
@yufenglee I am working in generate Qoperator format,but torch can only export QDQ format,is there any solution for QDQ to Qoperator ?
thanks

@JiliangNi
Copy link

Hello Yufeng, @yufenglee I am working in generate Qoperator format,but torch can only export QDQ format,is there any solution for QDQ to Qoperator ? thanks

@Rickustc Did you find any solution for QDQ to Qoperator? I met the similar problem and I found another quant tool which generates Qoperator model.

@UsingtcNower
Copy link

Hello Yufeng, @yufenglee I am working in generate Qoperator format,but torch can only export QDQ format,is there any solution for QDQ to Qoperator ? thanks

@Rickustc Did you find any solution for QDQ to Qoperator? I met the similar problem and I found another quant tool which generates Qoperator model.

I have the same question. @Rickustc @JiliangNi Did you find the solution?

@yufenglee
Copy link
Member

Please refer to #21137 for the answer.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

6 participants