dq_w  = dequant_mxfp4(layer.weight, layer.weight_scale, x.dtype)
qdq_x = self.quant_dequant_func(x)
return F.linear(qdq_x, dq_w, bias)