BUG: Duplicate bias addition in PseudoQuant4x16NoMasterFn causes incorrect inference output

[manyizhang]

In the file pseudoquant_linear_fns.py, inside the class PseudoQuant4x16NoMasterFn (line151-155), the forward pass contains the following code:

        y = torch.nn.functional.linear(x_flat_dq, weight_dq, bias)

        y = y.unflatten(dim=0, sizes=x.shape[:-1])
        if bias is not None:
            y += bias

The manual addition (y += bias) results in the bias being added twice, which leads to incorrect inference output when module has bias term .

[BlackSamorez]

You're right. Nobody ever tested biases so we missed it.

If you want to, you can just delete the y += bias lines where needed and open a PR. I'll merge it an push a new version.

[manyizhang]

Thank you for your prompting reply. I have opened a PR as mentioned. Please kindly check it. #18

You're right. Nobody ever tested biases so we missed it.

If you want to, you can just delete the y += bias lines where needed and open a PR. I'll merge it an push a new version.