Function make_quant_linear

inference/utils/load_quant.py:85–113

(module, names, w_bit, groupsize, device, name="")

def make_quant_linear(module, names, w_bit, groupsize, device, name=""):
    if isinstance(module, WQLinear):
        return
    for attr in dir(module):
        tmp = getattr(module, attr)
        name1 = name + "." + attr if name != "" else attr
        if name1 in names:
            delattr(module, attr)
            setattr(
                module,
                attr,
                WQLinear(
                    w_bit,
                    groupsize,
                    tmp.in_features,
                    tmp.out_features,
                    tmp.bias is not None,
                    device,
                ),
            )
    for name1, child in module.named_children():
        make_quant_linear(
            child,
            names,
            w_bit,
            groupsize,
            device,
            name + "." + name1 if name != "" else name1,
        )


def find_layers(module, layers=[nn.Linear], name=""):