Support torch QAT model export?

[WangFengtu1996]

Description of the Feature:

torch QAT supports three mode.

• Eager Mode Quantization
• FX Graph Mode Quantization
• PyTorch 2 Export Quantization detail in https://pytorch.org/docs/stable/quantization.html#module-torch.ao.quantization.backend_config.executorch

which mode is supprted by ai-edge-torch ? thkx

[pkgoogle]

Hi @WangFengtu1996, we are actively working on supporting more modes/recipes .. currently for PT Converter we support P2TE and original TFLite quantization: https://github.com/google-ai-edge/ai-edge-torch/blob/main/docs/pytorch_converter/README.md#quantization

For the generative API we support this: https://github.com/google-ai-edge/ai-edge-torch/tree/main/ai_edge_torch/generative/quantize

Does that answer your question?

[WangFengtu1996]

@pkgoogle
I need some support about Quantization Aware Training, not Post Training Quantization. Because the accuracy is too low by using Post Training Quantization.

[pkgoogle]

I am unsure if that is fully supported... I have found some resources which I think you can combine w/ the original TFLite Quantization above:

https://www.tensorflow.org/model_optimization/guide/quantization/training

You may need to change some steps to adjust for this workflow (which didn't exist at the time of the above documentation writing). Like you may want to switch to keras3 with PyTorch backend in the above steps and then attempt to use original TFLite Quantization.

[github-actions[bot]]

Marking this issue as stale since it has been open for 7 days with no activity. This issue will be closed if no further activity occurs.

[github-actions[bot]]

This issue was closed because it has been inactive for 14 days. Please post a new issue if you need further assistance. Thanks!