[tests] fix slow bart cnn test, faster marian tests (#7888)

tests/test_modeling_marian.py

@@ -16,7 +16,7 @@
 
 import unittest
 
-from transformers import is_torch_available
+from transformers import AutoConfig, AutoTokenizer, MarianConfig, MarianTokenizer, is_torch_available
 from transformers.file_utils import cached_property
 from transformers.hf_api import HfApi
 from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
@@ -25,14 +25,7 @@ from transformers.testing_utils import require_sentencepiece, require_tokenizers
 if is_torch_available():
     import torch
 
-    from transformers import (
-        AutoConfig,
-        AutoModelWithLMHead,
-        AutoTokenizer,
-        MarianConfig,
-        MarianMTModel,
-        MarianTokenizer,
-    )
+    from transformers import AutoModelWithLMHead, MarianMTModel
     from transformers.convert_marian_to_pytorch import (
         ORG_NAME,
         convert_hf_name_to_opus_name,
@@ -79,10 +72,16 @@ class MarianIntegrationTest(unittest.TestCase):
     @classmethod
     def setUpClass(cls) -> None:
         cls.model_name = f"Helsinki-NLP/opus-mt-{cls.src}-{cls.tgt}"
-        cls.tokenizer: MarianTokenizer = AutoTokenizer.from_pretrained(cls.model_name)
-        cls.eos_token_id = cls.tokenizer.eos_token_id
         return cls
 
+    @cached_property
+    def tokenizer(self) -> MarianTokenizer:
+        return AutoTokenizer.from_pretrained(self.model_name)
+
+    @property
+    def eos_token_id(self) -> int:
+        return self.tokenizer.eos_token_id
+
     @cached_property
     def model(self):
         model: MarianMTModel = AutoModelWithLMHead.from_pretrained(self.model_name).to(torch_device)