device agnostic models testing (#27146)

* device agnostic models testing

* add decorator `require_torch_fp16`

* make style

* apply review suggestion

* Oops, the fp16 decorator was misused

tests/models/mvp/test_modeling_mvp.py

@@ -22,7 +22,14 @@ import unittest
 import timeout_decorator  # noqa
 
 from transformers import MvpConfig, is_torch_available
-from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
+from transformers.testing_utils import (
+    require_sentencepiece,
+    require_tokenizers,
+    require_torch,
+    require_torch_fp16,
+    slow,
+    torch_device,
+)
 from transformers.utils import cached_property
 
 from ...generation.test_utils import GenerationTesterMixin
@@ -374,12 +381,12 @@ class MvpHeadTests(unittest.TestCase):
             mvp_toks = tokenizer.encode(ex, return_tensors="pt").squeeze()
             assert_tensors_close(desired_result.long(), mvp_toks, prefix=ex)
 
+    @require_torch_fp16
     def test_generate_fp16(self):
         config, input_ids, batch_size = self._get_config_and_data()
         attention_mask = input_ids.ne(1).to(torch_device)
         model = MvpForConditionalGeneration(config).eval().to(torch_device)
-        if torch_device == "cuda":
-            model.half()
+        model.half()
         model.generate(input_ids, attention_mask=attention_mask)
         model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
 
@@ -505,13 +512,13 @@ class MvpModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
             with torch.no_grad():
                 model(**inputs)[0]
 
+    @require_torch_fp16
     def test_generate_fp16(self):
         config, input_dict = self.model_tester.prepare_config_and_inputs()
         input_ids = input_dict["input_ids"]
         attention_mask = input_ids.ne(1).to(torch_device)
         model = MvpForConditionalGeneration(config).eval().to(torch_device)
-        if torch_device == "cuda":
-            model.half()
+        model.half()
         model.generate(input_ids, attention_mask=attention_mask)
         model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)