Fix gradient checkpointing bug in xglm (#22127)

src/transformers/models/xglm/modeling_xglm.py

@@ -714,6 +714,14 @@ class XGLMModel(XGLMPreTrainedModel):
 
         hidden_states = nn.functional.dropout(hidden_states, p=float(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
@@ -739,12 +747,6 @@ class XGLMModel(XGLMPreTrainedModel):
             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(
-                        "`use_cache = True` is incompatible with gradient checkpointing`. Setting `use_cache ="
-                        " False`..."
-                    )
-                    use_cache = False
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):