T5 & mT5 (#8552)

* add mt5 and t5v1_1 model

* fix tests

* correct some imports

* add tf model

* finish tf t5

* improve examples

* fix copies

* clean doc

tests/test_modeling_mt5.py

@@ -0,0 +1,39 @@
+import unittest
+
+from transformers import is_torch_available
+from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
+
+
+if is_torch_available():
+    from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
+
+
+@require_torch
+@require_sentencepiece
+@require_tokenizers
+class MT5IntegrationTest(unittest.TestCase):
+    @slow
+    def test_small_integration_test(self):
+        """
+        For comparision run:
+        >>> import t5  # pip install t5==0.7.1
+        >>> from t5.data.sentencepiece_vocabulary import SentencePieceVocabulary
+
+        >>> path_to_mtf_small_mt5_checkpoint = '<fill_in>'
+        >>> path_to_mtf_small_mt5_spm_model_path = '<fill_in>'
+        >>> t5_model = t5.models.MtfModel(model_dir=path_to_mtf_small_mt5_checkpoint, batch_size=1, tpu=None)
+        >>> vocab = SentencePieceVocabulary(path_to_mtf_small_mt5_spm_model_path)
+        >>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
+        """
+
+        model = AutoModelForSeq2SeqLM.from_pretrained("google/mt5-small", return_dict=True).to(torch_device)
+        tokenizer = AutoTokenizer.from_pretrained("google/mt5-small")
+
+        input_ids = tokenizer("Hello there", return_tensors="pt").input_ids
+        labels = tokenizer("Hi I am", return_tensors="pt").input_ids
+
+        loss = model(input_ids.to(torch_device), labels=labels.to(torch_device)).loss
+        mtf_score = -(labels.shape[-1] * loss.item())
+
+        EXPECTED_SCORE = -84.9127
+        self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)