Fix gradient checkpointing bug in trocr (#22126)

* Fix gradient checkpointing bug in trocr

* Fix format

* Update src/transformers/models/trocr/modeling_trocr.py
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>

---------
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>

src/transformers/models/trocr/modeling_trocr.py

@@ -664,6 +664,13 @@ class TrOCRDecoder(TrOCRPreTrainedModel):
             # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
             encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
 
+        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
@@ -689,12 +696,6 @@ class TrOCRDecoder(TrOCRPreTrainedModel):
             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):