Quantization method is different from paper?

[hustzxd]

Hello, thanks for your excellent work, and this project is beneficial to me.

However, I found that the quantization method in this project is somewhat different from that mentioned in the paper.

In the paper, the quantization function is:

Weights are quantized using:

which I think may be a mistake in the paper.

Activations are quantized using:

In this project, weights are quantized using:

Activations are quantized using: which I think may be a mistake.

In fact, according to quantize_module_.py, weights and activations are quantized using gemm method(an asymmetric uniform quantization method)

Is there any comparison between the two methods?

Best wishes.

[bohanzhuang]

There is a small mistake in weights quantization in the paper. His weights quantization implementation is correct.