fix(mge/optimizer): only disable convert inputs in build-in optimizers

GitOrigin-RevId: 1a48fe318dc29c80c8f244923d64de79d9abd6b8

imperative/python/megengine/optimizer/adadelta.py

@@ -48,6 +48,7 @@ class Adadelta(Optimizer):
 
         defaults = dict(lr=lr, rho=rho, eps=eps, weight_decay=weight_decay)
         super().__init__(params, defaults)
+        self._disable_type_convert = True
 
     def _create_state(self, param_group):
         for param in param_group["params"]:

imperative/python/megengine/optimizer/adagrad.py

@@ -48,6 +48,7 @@ class Adagrad(Optimizer):
 
         defaults = dict(lr=lr, lr_decay=lr_decay, eps=eps, weight_decay=weight_decay)
         super().__init__(params, defaults)
+        self._disable_type_convert = True
 
     def _create_state(self, param_group):
         for param in param_group["params"]:

imperative/python/megengine/optimizer/adam.py

@@ -47,6 +47,7 @@ class Adam(Optimizer):
 
         defaults = dict(lr=lr, weight_decay=weight_decay, betas=betas, eps=eps)
         super().__init__(params, defaults)
+        self._disable_type_convert = True
 
     def _create_state(self, param_group):
         for param in param_group["params"]:

imperative/python/megengine/optimizer/adamw.py

@@ -47,6 +47,7 @@ class AdamW(Optimizer):
 
         defaults = dict(lr=lr, weight_decay=weight_decay, betas=betas, eps=eps)
         super().__init__(params, defaults)
+        self._disable_type_convert = True
 
     def _create_state(self, param_group):
         for param in param_group["params"]:

imperative/python/megengine/optimizer/optimizer.py

@@ -42,6 +42,7 @@ class Optimizer(metaclass=ABCMeta):
     ):
         self._state = dict()
         self._defaults = defaults
+        self._disable_type_convert = False
 
         if isinstance(params, (Parameter, dict)):
             params = [params]
@@ -149,6 +150,7 @@ class Optimizer(metaclass=ABCMeta):
         # set the globle state `_enable_convert_inputs` to `False` to disable
         # the `convert_inputs` for param updates
         set_option("record_computing_path", 0)
+        if self._disable_type_convert:
             backup = set_convert_inputs(False)
         for group in self.param_groups:
             if isinstance(group["params"], set):
@@ -160,6 +162,7 @@ class Optimizer(metaclass=ABCMeta):
             push_scope("step")
             self._updates(group)
             pop_scope("step")
+        if self._disable_type_convert:
             # restore the globle state `_enable_convert_inputs`
             set_convert_inputs(backup)
         set_option("record_computing_path", 1)

imperative/python/megengine/optimizer/sgd.py

@@ -43,6 +43,7 @@ class SGD(Optimizer):
 
         defaults = dict(lr=lr, momentum=momentum, weight_decay=weight_decay)
         super().__init__(params, defaults)
+        self._disable_type_convert = True
 
     def _create_state(self, param_group):
         if param_group["momentum"] != 0.0: