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/plbart/test_modeling_plbart.py

@@ -496,7 +496,7 @@ class PLBartStandaloneDecoderModelTester:
         decoder_layers=2,
         encoder_attention_heads=4,
         decoder_attention_heads=4,
-        max_position_embeddings=30,
+        max_position_embeddings=50,
         is_encoder_decoder=False,
         pad_token_id=0,
         bos_token_id=1,
@@ -634,9 +634,9 @@ class PLBartStandaloneDecoderModelTester:
 
         # get two different outputs
         output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
-        output_from_past = model(next_tokens, attention_mask=attn_mask, past_key_values=past_key_values)[
-            "last_hidden_state"
-        ]
+        output_from_past = model(
+            next_tokens, attention_mask=attn_mask, past_key_values=past_key_values, use_cache=True
+        )["last_hidden_state"]
 
         # select random slice
         random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()