Fix gradient checkpointing bug in Rembert (#21945)

src/transformers/models/rembert/modeling_rembert.py

@@ -521,6 +521,12 @@ class RemBertEncoder(nn.Module):
         output_hidden_states: bool = False,
         return_dict: bool = True,
     ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
         hidden_states = self.embedding_hidden_mapping_in(hidden_states)
         all_hidden_states = () if output_hidden_states else None
         all_self_attentions = () if output_attentions else None
@@ -535,11 +541,6 @@ class RemBertEncoder(nn.Module):
             past_key_value = past_key_values[i] if past_key_values is not None else None
 
             if self.gradient_checkpointing and self.training:
-                if use_cache:
-                    logger.warning_once(
-                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                    )
-                    use_cache = False
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):