Fix gradient checkpointing megatron bert (#21921)

src/transformers/models/megatron_bert/modeling_megatron_bert.py

@@ -530,6 +530,12 @@ class MegatronBertEncoder(nn.Module):
         output_hidden_states: Optional[bool] = False,
         return_dict: Optional[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
         all_hidden_states = () if output_hidden_states else None
         all_self_attentions = () if output_attentions else None
         all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
@@ -543,11 +549,6 @@ class MegatronBertEncoder(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):