where the activation/feature-map is quantized?

[xieyi4650]

❓ Questions

(Please ask your question here.)
Hi after studying code, I found that the UniformQuantizer quantize the layer weights using forward pre hook
But what confused me is that, I have not found where is the quantize operation on activation.
seems that the hook do not operate the input
Thanks

[adefossez]

Hello @xieyi4650
As mentioned in the paper, we performed the activation quantization experiments only on the Transformer on Wikitext task, using PyTorch quantization framework post training. You can find the code to perform that operation following the instructions and README of that example.

We did not perform any activation quantization experiments for CIFAR models, be it for the QAT models, or the DiffQ models. While it would be interesting to experiment with a DiffQ-like approach for activation quantization, we left that for future work.

Cheers