t5 model should make decoder_attention_mask (#6800)

src/transformers/tokenization_t5.py

@@ -372,6 +372,5 @@ class T5Tokenizer(PreTrainedTokenizer):
             **kwargs,
         )
         model_inputs["labels"] = labels_and_decoder_mask["input_ids"]
-        model_inputs["decoder_attention_mask"] = labels_and_decoder_mask["attention_mask"]
         self.prefix_tokens = []
         return model_inputs

tests/test_tokenization_t5.py

@@ -162,14 +162,12 @@ class T5TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
             src_text, tgt_texts=tgt_text, max_target_length=32, padding="max_length", return_tensors=FRAMEWORK
         )
         self.assertEqual(32, batch["labels"].shape[1])
-        self.assertEqual(32, batch["decoder_attention_mask"].shape[1])
 
         # test None max_target_length
         batch = tokenizer.prepare_seq2seq_batch(
             src_text, tgt_texts=tgt_text, max_length=32, padding="max_length", return_tensors=FRAMEWORK
         )
         self.assertEqual(32, batch["labels"].shape[1])
-        self.assertEqual(32, batch["decoder_attention_mask"].shape[1])
 
     def test_outputs_not_longer_than_maxlen(self):
         tokenizer = self.t5_base_tokenizer