TFMarian, TFMbart, TFPegasus, TFBlenderbot (#7987)

* Start plumbing

* Marian close

* Small stubs for all children

* Fixed bart

* marian working

* pegasus test is good, but failing

* Checkin tests

* More model files

* Subtle marian, pegasus integration test failures

* Works well

* rm print

* boom boom

* Still failing model2doc

* merge master

* Equivalence test failing, all others fixed

* cleanup

* Fix embed_scale

* Cleanup marian pipeline test

* Undo extra changes

* Smaller delta

* Cleanup model testers

* undo delta

* fix tests import structure

* cross test decorator

* Cleaner set_weights

* Respect authorized_unexpected_keys

* No warnings

* No warnings

* style

* Nest tf import

* black

* Apply suggestions from code review

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* functional dropout

* fixup

* Fixup

* style_doc

* embs

* shape list

* delete slow force_token_id_to_be_generated func

* fixup

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

tests/test_modeling_marian.py

@@ -138,7 +138,7 @@ class MarianIntegrationTest(unittest.TestCase):
         )
         self.assertEqual(self.model.device, model_inputs.input_ids.device)
         generated_ids = self.model.generate(
-            model_inputs.input_ids, attention_mask=model_inputs.attention_mask, num_beams=2
+            model_inputs.input_ids, attention_mask=model_inputs.attention_mask, num_beams=2, max_length=128
         )
         generated_words = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
         return generated_words
@@ -244,6 +244,8 @@ class TestMarian_RU_FR(MarianIntegrationTest):
 @require_sentencepiece
 @require_tokenizers
 class TestMarian_MT_EN(MarianIntegrationTest):
+    """Cover low resource/high perplexity setting. This breaks without adjust_logits_generation overwritten"""
+
     src = "mt"
     tgt = "en"
     src_text = ["Billi messu b'mod ġentili, Ġesù fejjaq raġel li kien milqut bil - marda kerha tal - ġdiem."]