diff --git a/paddleslim/quant/advanced/layerwise_quant_error.py b/paddleslim/quant/advanced/layerwise_quant_error.py
index ce03d198cd09a943cedde26b0338ff50f65b8e12..ae230ceddfa30ae1653ca5f7b92e886c8919a3ed 100644
--- a/paddleslim/quant/advanced/layerwise_quant_error.py
+++ b/paddleslim/quant/advanced/layerwise_quant_error.py
@@ -54,6 +54,7 @@ class LayerWiseQuantError(nn.Layer):
             if type(cur_layer) == LayerWiseQuantError:
                 print(cur_name, cur_layer.losses.mean())
         '''
+        super(LayerWiseQuantError, self).__init__()
         self.layer = layer
         self.weight = layer.weight
         self.weight_bits = weight_bits
@@ -62,14 +63,13 @@ class LayerWiseQuantError(nn.Layer):
         self.act_method = act_quant_method
         self.loss_function = loss_function
         self.losses = []
+        self.loss = None
 
     def forward(self, input):
         act = input[0] if type(input) == tuple else input
         origin_out = paddle.matmul(act, self.weight)
         bnt = (1 << (self.weight_bits - 1)) - 1
-        quant_scale = compute_scales(
-            self.weight.cast('float32'),
-            method=self.weight_method).cast(self.weight.dtype)
+        quant_scale = compute_scales(self.weight, method=self.weight_method)
         quant_weight = paddle.clip(
             paddle.round(self.weight / quant_scale * bnt), -bnt - 1, bnt)
         quant_dequant_weight = quant_weight / bnt * quant_scale
@@ -80,6 +80,7 @@ class LayerWiseQuantError(nn.Layer):
             paddle.round(act / quant_scale * bnt), -bnt - 1, bnt)
         quant_dequant_act = quant_act / bnt * quant_scale
         quant_out = paddle.matmul(quant_dequant_act, quant_dequant_weight)
-        loss = self.loss_function(origin_out, quant_out)
+        loss = self.loss_function(origin_out, quant_out).cast('float32')
         self.losses.append(loss)
+        self.loss = paddle.to_tensor(self.losses, dtype='float32').mean()
         return self.layer(input)
diff --git a/paddleslim/quant/advanced/utils.py b/paddleslim/quant/advanced/utils.py
index 98f24ef124e10b345be6354f9c37fc352cfee6de..703fc5e1ce6b6e2f7419ab6c494c51d4efd5f52b 100644
--- a/paddleslim/quant/advanced/utils.py
+++ b/paddleslim/quant/advanced/utils.py
@@ -48,8 +48,10 @@ def compute_scales(x, method='abs_max'):
     elif method == 'abs_max_channel_wise':
         reduce_axis = tuple([i for i in range(len(x.shape)) if i != 1])
         quant_scale = paddle.max(paddle.abs(x), axis=reduce_axis)
-        quant_scale = paddle.where(quant_scale == np.float32(0.0),
-                                   np.float32(1e-8), quant_scale)
+        quant_scale = paddle.where(quant_scale == paddle.to_tensor(
+            0, dtype=x.dtype),
+                                   paddle.to_tensor(1e-8, dtype=x.dtype),
+                                   quant_scale)
     return quant_scale
 
 
diff --git a/paddleslim/quant/advanced/utils_layers.py b/paddleslim/quant/advanced/utils_layers.py
index 3b06a3ede83a3ef3892026d2ef87c0284e8c320c..166d938da2cadfa27e42ea5621b78f5d02f0ea16 100644
--- a/paddleslim/quant/advanced/utils_layers.py
+++ b/paddleslim/quant/advanced/utils_layers.py
@@ -26,7 +26,7 @@ class ShiftSmoothHelpLayer(nn.Layer):
         super(ShiftSmoothHelpLayer, self).__init__()
         self.weight = layer.weight
         shift_shape = self.weight.shape[0]
-        if hasattr(layer, "bias") or layer.bias is None:
+        if not hasattr(layer, "bias") or layer.bias is None:
             self.bias = paddle.create_parameter(
                 shape=[self.weight.shape[1]],
                 dtype=self.weight.dtype,