Fix gradient checkpointing bug M2M 100 (#21841)

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/models/m2m_100/modeling_m2m_100.py

@@ -1024,6 +1024,13 @@ class M2M100Decoder(M2M100PreTrainedModel):
 
         hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
 
+        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
+
         # decoder layers
         all_hidden_states = () if output_hidden_states else None
         all_self_attns = () if output_attentions else None
@@ -1054,12 +1061,6 @@ class M2M100Decoder(M2M100PreTrainedModel):
                 past_key_value = past_key_values[idx] 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):