diff --git a/src/transformers/models/rembert/modeling_tf_rembert.py b/src/transformers/models/rembert/modeling_tf_rembert.py
index 16accc1b944..9bf6ba6edee 100644
--- a/src/transformers/models/rembert/modeling_tf_rembert.py
+++ b/src/transformers/models/rembert/modeling_tf_rembert.py
@@ -477,7 +477,7 @@ class TFRemBertEncoder(tf.keras.layers.Layer):
         training: bool = False,
     ) -> Union[TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor]]:
         hidden_states = self.embedding_hidden_mapping_in(inputs=hidden_states)
-        all_hidden_states = (hidden_states,) if output_hidden_states else None
+        all_hidden_states = () if output_hidden_states else None
         all_attentions = () if output_attentions else None
         all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
 
diff --git a/tests/test_modeling_tf_rembert.py b/tests/test_modeling_tf_rembert.py
index 8908e6d02bd..a6d7d35cf98 100644
--- a/tests/test_modeling_tf_rembert.py
+++ b/tests/test_modeling_tf_rembert.py
@@ -90,9 +90,6 @@ class TFRemBertModelTester:
         self.num_choices = 4
         self.scope = None
 
-        # RemBERT also returns the upprojected word embeddings as an hidden layers
-        self.expected_num_hidden_layers = self.num_hidden_layers + 2
-
     def prepare_config_and_inputs(self):
         input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)