use return dict for rag encoder (#9363)

src/transformers/models/rag/modeling_rag.py

@@ -1437,7 +1437,7 @@ class RagTokenForGeneration(RagPreTrainedModel):
         batch_size = context_input_ids.shape[0] // n_docs
 
         encoder = self.rag.generator.get_encoder()
-        encoder_outputs = encoder(input_ids=context_input_ids, attention_mask=context_attention_mask)
+        encoder_outputs = encoder(input_ids=context_input_ids, attention_mask=context_attention_mask, return_dict=True)
 
         input_ids = torch.full(
             (batch_size * num_beams, 1),