Bart: new cache format (#35314)

* bart compile

* add mbart

* some more models touched by fix-copies

* more

* more models

* even more models

* fix copies

* fix tests

* fix copies

* fix

* biogpt accepts position ids now (breaking?)

* fix failing non-slow tests

* fix some tests

* should not be removed

* small update

* Update src/transformers/models/bart/modeling_bart.py

Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* update for last `main`

* fix copies

* clone `update_causal_mask` from llama

* tmp

* fixup

* why? how?

* fix bart tests

* dont skip test

* address comments

* fix tests

* fix

* fixup and delete the file

---------

Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

tests/models/m2m_100/test_modeling_m2m_100.py

@@ -82,7 +82,7 @@ class M2M100ModelTester:
         attention_probs_dropout_prob=0.1,
         encoder_layerdrop=0.0,
         decoder_layerdrop=0.0,
-        max_position_embeddings=20,
+        max_position_embeddings=50,
         eos_token_id=2,
         pad_token_id=1,
         bos_token_id=0,
@@ -426,7 +426,7 @@ class M2M100ModelIntegrationTests(unittest.TestCase):
         Overwriting the common test as the test is flaky on tiny models
         """
         model = M2M100ForConditionalGeneration.from_pretrained(
-            "facebook/m2m100_418M", attn_implementation="flash_attention_2"
+            "facebook/m2m100_418M", torch_dtype=torch.float16, attn_implementation="flash_attention_2"
         ).to(torch_device)
 
         tokenizer = M2M100Tokenizer.from_pretrained("facebook/m2m100_418M", src_lang="fr", tgt_lang="en")