fix(mge/quant): avoid updating the scale and zero point in eval mode

GitOrigin-RevId: dfb08cf701

python_module/megengine/module/module.py

@@ -294,7 +294,8 @@ class Module(metaclass=ABCMeta):
         self.training = mode
 
         def fn(x) -> None:
-            x.training = mode
+            if x is not self:
+                x.train(mode=mode)
 
         self.apply(fn)
 

python_module/megengine/quantization/observer.py

@@ -56,6 +56,13 @@ class Observer(Module):
     def disable(self):
         self.enabled = False
 
+    def train(self, mode: bool = True) -> None:
+        super().train(mode)
+        if mode:
+            self.enable()
+        else:
+            self.disable()
+
     @abstractmethod
     def forward(self, x):
         pass