diff --git a/tf_adapter/python/npu_bridge/embedding/embedding_optimizer.py b/tf_adapter/python/npu_bridge/embedding/embedding_optimizer.py
index 3ee6aef85b9f47032abbe448e845aba10ca76639..554382b6de9b6229e18da5f907f04c0996026b0b 100644
--- a/tf_adapter/python/npu_bridge/embedding/embedding_optimizer.py
+++ b/tf_adapter/python/npu_bridge/embedding/embedding_optimizer.py
@@ -48,6 +48,9 @@ class AdamOptimizer(adam.AdamOptimizer):
         super(AdamOptimizer, self).__init__(learning_rate, beta_1, beta_2, epsilon, using_locking, name)
         self._beta1_power = None
         self._beta2_power = None
+        self._table_index = 0
+        self.max_key_num_list = list()
+        self.embedding_dim_list = list()
 
     @property
     def embedding_dims(self):
@@ -100,9 +103,12 @@ class AdamOptimizer(adam.AdamOptimizer):
                                                           grad,
                                                           indices,
                                                           ops.convert_to_tensor(_GLOBAL_STEP_VALUE),
-                                                          self._embedding_dims)
-            result.op._set_attr("_embedding_dim", attr_value_pb2.AttrValue(i=self._embedding_dims))
-            result.op._set_attr("_max_key_num", attr_value_pb2.AttrValue(i=self._max_nums))
+                                                          self.embedding_dim_list[self._table_index])
+            result.op._set_attr("_embedding_dim",
+                                attr_value_pb2.AttrValue(i=self.embedding_dim_list[self._table_index]))
+            result.op._set_attr("_max_key_num",
+                                attr_value_pb2.AttrValue(i=self.max_key_num_list[self._table_index]))
+            self._table_index += 1
             return result
         else:
             return self._apply_sparse_shared(grad, var, indices, self._resource_scatter_add)
@@ -119,6 +125,17 @@ class AdamOptimizer(adam.AdamOptimizer):
 
 
 class AdagradOptimizer(adagrad.AdagradOptimizer):
+    def __init__(self,
+                 learning_rate,
+                 initial_accumulator_value=0.1,
+                 use_locking=False,
+                 name="EmbeddingAdagradOptimizer"):
+        """Construct a AdamW optimizer."""
+        super(AdagradOptimizer, self).__init__(learning_rate, initial_accumulator_value, use_locking, name)
+        self._table_index = 0
+        self.max_key_num_list = list()
+        self.embedding_dim_list = list()
+
     @property
     def embedding_dims(self):
         return self._embedding_dims
@@ -150,9 +167,12 @@ class AdagradOptimizer(adagrad.AdagradOptimizer):
                                                               grad,
                                                               indices,
                                                               ops.convert_to_tensor(_GLOBAL_STEP_VALUE),
-                                                              self._embedding_dims)
-            result.op._set_attr("_embedding_dim", attr_value_pb2.AttrValue(i=self._embedding_dims))
-            result.op._set_attr("_max_key_num", attr_value_pb2.AttrValue(i=self._max_nums))
+                                                              self.embedding_dim_list[self._table_index])
+            result.op._set_attr("_embedding_dim",
+                                attr_value_pb2.AttrValue(i=self.embedding_dim_list[self._table_index]))
+            result.op._set_attr("_max_key_num",
+                                attr_value_pb2.AttrValue(i=self.max_key_num_list[self._table_index]))
+            self._table_index += 1
             return result
         else:
             return self.training_ops.resource_sparse_apply_adagrad(var.handle, grad.handle,
@@ -213,6 +233,9 @@ class AdamWOptimizer(optimizer.Optimizer):
         self._max_grad_norm_t = None
         self._beta1_power_t = None
         self._beta2_power_t = None
+        self._table_index = 0
+        self.max_key_num_list = list()
+        self.embedding_dim_list = list()
 
     @property
     def embedding_dims(self):
@@ -276,9 +299,12 @@ class AdamWOptimizer(optimizer.Optimizer):
                                                             math_ops.cast(self._max_grad_norm_t, grad.dtype),
                                                             amsgrad=self._amsgrad,
                                                             maximize=self._maximize,
-                                                            embedding_dim=self._embedding_dims)
-            result.op._set_attr("_embedding_dim", attr_value_pb2.AttrValue(i=self._embedding_dims))
-            result.op._set_attr("_max_key_num", attr_value_pb2.AttrValue(i=self._max_nums))
+                                                            embedding_dim=self.embedding_dim_list[self._table_index])
+            result.op._set_attr("_embedding_dim",
+                                attr_value_pb2.AttrValue(i=self.embedding_dim_list[self._table_index]))
+            result.op._set_attr("_max_key_num",
+                                attr_value_pb2.AttrValue(i=self.max_key_num_list[self._table_index]))
+            self._table_index += 1
             return result
         else:
             raise TypeError("Variable is not NpuEmbeddingResource type, please check.")
diff --git a/tf_adapter/python/npu_bridge/embedding/embedding_service.py b/tf_adapter/python/npu_bridge/embedding/embedding_service.py
index 9540613b9065e697e304ce539901bfaf9c2f5df7..486f7db71996cbce2e947fdd3f15fd2bc5917492 100644
--- a/tf_adapter/python/npu_bridge/embedding/embedding_service.py
+++ b/tf_adapter/python/npu_bridge/embedding/embedding_service.py
@@ -233,8 +233,8 @@ class ESWorker:
                     (initializer != 'truncated_normal') and (initializer != 'constant'):
                 raise ValueError("initializer must be random_uniform or truncated_normal or constant.")
             self._optimizer = optimizer
-            self._optimizer._embedding_dims = embedding_dim
-            self._optimizer._max_nums = max_batch_size
+            self._optimizer.embedding_dim_list.append(embedding_dim)
+            self._optimizer.max_key_num_list.append(max_batch_size)
             self._optimizer._es_cluster_configs = self._es_cluster_conf
             self._table_to_optimizer[table_id] = self._optimizer
             # adam include m and v, 2 slots; adagrad include accumulator, 1 slot