diff --git a/imperative/python/megengine/quantization/fake_quant.py b/imperative/python/megengine/quantization/fake_quant.py
index c1469d33fbd86fe8e4bf0c388270cb4b4c318d00..f6625d7083e063893b532e40218a6942798055e5 100644
--- a/imperative/python/megengine/quantization/fake_quant.py
+++ b/imperative/python/megengine/quantization/fake_quant.py
@@ -132,7 +132,7 @@ class LSQ(_FakeQuantize, QParamsModuleMixin):
     :param eps:a small value to avoid division by zero. Default: 1e-5
     """
 
-    def init(
+    def __init__(
         self,
         dtype: Union[str, QuantDtypeMeta],
         enable: bool = True,
@@ -142,6 +142,9 @@ class LSQ(_FakeQuantize, QParamsModuleMixin):
         super().__init__(dtype=dtype, enable=enable, **kwargs)
         self.eps = Tensor(eps, dtype="float32")
         self.step_size = Parameter(1.0, dtype="float32")
+        self.mode = None
+        self.zero_point = Tensor(0.0, dtype="float32")
+        self.grad_scale = Tensor(1.0, dtype="float32")
 
     def set_qparams(self, qparams: LSQParams):
         self.mode = qparams.mode