Fix slow gpu tests lysandre (#4487)

* There is one missing key in BERT

* Correct device for CamemBERT model

* RoBERTa tokenization adding prefix space

* Style

tests/test_modeling_auto.py

@@ -80,8 +80,9 @@ class AutoModelTest(unittest.TestCase):
             model, loading_info = AutoModelForPreTraining.from_pretrained(model_name, output_loading_info=True)
             self.assertIsNotNone(model)
             self.assertIsInstance(model, BertForPreTraining)
-            for value in loading_info.values():
-                self.assertEqual(len(value), 0)
+            for key, value in loading_info.items():
+                # Only one value should not be initialized and in the missing keys.
+                self.assertEqual(len(value), 1 if key == "missing_keys" else 0)
 
     @slow
     def test_lmhead_model_from_pretrained(self):

tests/test_modeling_camembert.py

@@ -30,6 +30,7 @@ class CamembertModelIntegrationTest(unittest.TestCase):
     @slow
     def test_output_embeds_base_model(self):
         model = CamembertModel.from_pretrained("camembert-base")
+        model.to(torch_device)
 
         input_ids = torch.tensor(
             [[5, 121, 11, 660, 16, 730, 25543, 110, 83, 6]], device=torch_device, dtype=torch.long,

tests/test_tokenization_roberta.py

@@ -100,9 +100,11 @@ class RobertaTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
         text = tokenizer.encode("sequence builders", add_special_tokens=False)
         text_2 = tokenizer.encode("multi-sequence build", add_special_tokens=False)
 
-        encoded_text_from_decode = tokenizer.encode("sequence builders", add_special_tokens=True)
+        encoded_text_from_decode = tokenizer.encode(
+            "sequence builders", add_special_tokens=True, add_prefix_space=False
+        )
         encoded_pair_from_decode = tokenizer.encode(
-            "sequence builders", "multi-sequence build", add_special_tokens=True
+            "sequence builders", "multi-sequence build", add_special_tokens=True, add_prefix_space=False
         )
 
         encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)