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/generation/test_utils.py

@@ -22,7 +22,13 @@ import warnings
 import numpy as np
 
 from transformers import is_torch_available, pipeline
-from transformers.testing_utils import require_accelerate, require_torch, require_torch_multi_gpu, slow, torch_device
+from transformers.testing_utils import (
+    require_accelerate,
+    require_torch,
+    require_torch_multi_accelerator,
+    slow,
+    torch_device,
+)
 
 from ..test_modeling_common import floats_tensor, ids_tensor
 from .test_framework_agnostic import GenerationIntegrationTestsMixin
@@ -1019,7 +1025,7 @@ class GenerationTesterMixin:
                 )
 
     @require_accelerate
-    @require_torch_multi_gpu
+    @require_torch_multi_accelerator
     def test_model_parallel_beam_search(self):
         for model_class in self.all_generative_model_classes:
             if model_class._no_split_modules is None:

tests/models/bark/test_modeling_bark.py

@@ -32,7 +32,13 @@ from transformers.models.bark.generation_configuration_bark import (
     BarkFineGenerationConfig,
     BarkSemanticGenerationConfig,
 )
-from transformers.testing_utils import require_torch, require_torch_gpu, slow, torch_device
+from transformers.testing_utils import (
+    require_torch,
+    require_torch_fp16,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
 from transformers.utils import cached_property
 
 from ...generation.test_utils import GenerationTesterMixin
@@ -570,12 +576,12 @@ class BarkSemanticModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.Te
             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 = self.all_generative_model_classes[0](config).eval().to(torch_device)
-        if torch_device == "cuda":
         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)
@@ -636,12 +642,12 @@ class BarkCoarseModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.Test
             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 = self.all_generative_model_classes[0](config).eval().to(torch_device)
-        if torch_device == "cuda":
         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)
@@ -700,13 +706,13 @@ class BarkFineModelTest(ModelTesterMixin, unittest.TestCase):
             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"]
         # take first codebook channel
 
         model = self.all_model_classes[0](config).eval().to(torch_device)
-        if torch_device == "cuda":
         model.half()
 
         # toy generation_configs

tests/models/bart/test_modeling_bart.py

@@ -22,7 +22,14 @@ import unittest
 import timeout_decorator  # noqa
 
 from transformers import BartConfig, 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
@@ -383,11 +390,11 @@ class BartHeadTests(unittest.TestCase):
             bart_toks = tokenizer.encode(ex, return_tensors="pt").squeeze()
             assert_tensors_close(desired_result.long(), bart_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 = BartForConditionalGeneration(config).eval().to(torch_device)
-        if torch_device == "cuda":
         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)
@@ -497,12 +504,12 @@ class BartModelTest(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 = BartForConditionalGeneration(config).eval().to(torch_device)
-        if torch_device == "cuda":
         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)

tests/models/bert/test_modeling_bert.py

@@ -18,7 +18,7 @@ import unittest
 
 from transformers import BertConfig, is_torch_available
 from transformers.models.auto import get_values
-from transformers.testing_utils import CaptureLogger, require_torch, require_torch_gpu, slow, torch_device
+from transformers.testing_utils import CaptureLogger, require_torch, require_torch_accelerator, slow, torch_device
 
 from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
@@ -601,7 +601,7 @@ class BertModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin
             self.assertIsNotNone(model)
 
     @slow
-    @require_torch_gpu
+    @require_torch_accelerator
     def test_torchscript_device_change(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
         for model_class in self.all_model_classes:

tests/models/bigbird_pegasus/test_modeling_bigbird_pegasus.py

@@ -20,7 +20,14 @@ import tempfile
 import unittest
 
 from transformers import BigBirdPegasusConfig, 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 ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
@@ -376,12 +383,12 @@ class BigBirdPegasusModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineT
             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_dict.pop("decoder_attention_mask")
         input_dict.pop("decoder_input_ids")
         model = BigBirdPegasusForConditionalGeneration(config).eval().to(torch_device)
-        if torch_device == "cuda":
         model.half()
         model.generate(**input_dict)
         model.generate(**input_dict, do_sample=True, early_stopping=False, num_return_sequences=3)

tests/models/blenderbot/test_modeling_blenderbot.py

@@ -18,7 +18,15 @@ import tempfile
 import unittest
 
 from transformers import BlenderbotConfig, is_torch_available
-from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
+from transformers.testing_utils import (
+    backend_empty_cache,
+    require_sentencepiece,
+    require_tokenizers,
+    require_torch,
+    require_torch_fp16,
+    slow,
+    torch_device,
+)
 from transformers.utils import cached_property
 
 from ...generation.test_utils import GenerationTesterMixin
@@ -269,12 +277,12 @@ class BlenderbotModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTeste
         config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
         self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs)
 
+    @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 = BlenderbotForConditionalGeneration(config).eval().to(torch_device)
-        if torch_device == "cuda":
         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)
@@ -315,7 +323,7 @@ class Blenderbot3BIntegrationTests(unittest.TestCase):
         FASTER_GEN_KWARGS = {"num_beams": 1, "early_stopping": True, "min_length": 15, "max_length": 25}
         TOK_DECODE_KW = {"skip_special_tokens": True, "clean_up_tokenization_spaces": True}
 
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
         model = BlenderbotForConditionalGeneration.from_pretrained(self.ckpt).half().to(torch_device)
 
         src_text = ["Sam"]

tests/models/blenderbot_small/test_modeling_blenderbot_small.py

@@ -18,7 +18,12 @@ import tempfile
 import unittest
 
 from transformers import BlenderbotSmallConfig, is_torch_available
-from transformers.testing_utils import require_torch, slow, torch_device
+from transformers.testing_utils import (
+    require_torch,
+    require_torch_fp16,
+    slow,
+    torch_device,
+)
 from transformers.utils import cached_property
 
 from ...generation.test_utils import GenerationTesterMixin
@@ -269,12 +274,12 @@ class BlenderbotSmallModelTest(ModelTesterMixin, GenerationTesterMixin, Pipeline
         config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
         self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs)
 
+    @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 = BlenderbotSmallForConditionalGeneration(config).eval().to(torch_device)
-        if torch_device == "cuda":
         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)

tests/models/blip/test_modeling_blip.py

@@ -24,7 +24,14 @@ import numpy as np
 import requests
 
 from transformers import BlipConfig, BlipTextConfig, BlipVisionConfig
-from transformers.testing_utils import require_torch, require_torch_gpu, require_vision, slow, torch_device
+from transformers.testing_utils import (
+    require_torch,
+    require_torch_accelerator,
+    require_torch_fp16,
+    require_vision,
+    slow,
+    torch_device,
+)
 from transformers.utils import is_torch_available, is_vision_available
 
 from ...test_configuration_common import ConfigTester
@@ -1280,7 +1287,8 @@ class BlipModelIntegrationTest(unittest.TestCase):
             [30522, 1037, 3861, 1997, 1037, 2450, 1998, 2014, 3899, 2006, 1996, 3509, 102],
         )
 
-    @require_torch_gpu
+    @require_torch_accelerator
+    @require_torch_fp16
     def test_inference_image_captioning_fp16(self):
         model = BlipForConditionalGeneration.from_pretrained(
             "Salesforce/blip-image-captioning-base", torch_dtype=torch.float16

tests/models/blip_2/test_modeling_blip_2.py

@@ -23,7 +23,13 @@ import numpy as np
 import requests
 
 from transformers import CONFIG_MAPPING, Blip2Config, Blip2QFormerConfig, Blip2VisionConfig
-from transformers.testing_utils import require_torch, require_torch_multi_gpu, require_vision, slow, torch_device
+from transformers.testing_utils import (
+    require_torch,
+    require_torch_multi_accelerator,
+    require_vision,
+    slow,
+    torch_device,
+)
 from transformers.utils import is_torch_available, is_vision_available
 
 from ...test_configuration_common import ConfigTester
@@ -934,8 +940,8 @@ class Blip2ModelIntegrationTest(unittest.TestCase):
         self.assertEqual(predictions[0].tolist(), [0, 2335, 1556, 28, 1782, 30, 8, 2608, 1])
         self.assertEqual(predictions[1].tolist(), [0, 2335, 1556, 28, 1782, 30, 8, 2608, 1])
 
-    @require_torch_multi_gpu
-    def test_inference_opt_multi_gpu(self):
+    @require_torch_multi_accelerator
+    def test_inference_opt_multi_accelerator(self):
         processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
         model = Blip2ForConditionalGeneration.from_pretrained(
             "Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16, device_map="balanced"
@@ -966,8 +972,8 @@ class Blip2ModelIntegrationTest(unittest.TestCase):
         )
         self.assertEqual(generated_text, "it's not a city, it's a beach")
 
-    @require_torch_multi_gpu
-    def test_inference_t5_multi_gpu(self):
+    @require_torch_multi_accelerator
+    def test_inference_t5_multi_accelerator(self):
         processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xl")
         device_map = device_map = {
             "query_tokens": 0,

tests/models/convbert/test_modeling_convbert.py

@@ -19,7 +19,7 @@ import unittest
 
 from transformers import ConvBertConfig, is_torch_available
 from transformers.models.auto import get_values
-from transformers.testing_utils import require_torch, require_torch_gpu, slow, torch_device
+from transformers.testing_utils import require_torch, require_torch_accelerator, slow, torch_device
 
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
@@ -427,7 +427,7 @@ class ConvBertModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase
                 )
 
     @slow
-    @require_torch_gpu
+    @require_torch_accelerator
     def test_torchscript_device_change(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
         for model_class in self.all_model_classes:

tests/models/ctrl/test_modeling_ctrl.py

@@ -17,7 +17,7 @@ import gc
 import unittest
 
 from transformers import CTRLConfig, is_torch_available
-from transformers.testing_utils import require_torch, slow, torch_device
+from transformers.testing_utils import backend_empty_cache, require_torch, slow, torch_device
 
 from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
@@ -230,7 +230,7 @@ class CTRLModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin
         super().tearDown()
         # clean-up as much as possible GPU memory occupied by PyTorch
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def test_config(self):
         self.config_tester.run_common_tests()
@@ -260,7 +260,7 @@ class CTRLModelLanguageGenerationTest(unittest.TestCase):
         super().tearDown()
         # clean-up as much as possible GPU memory occupied by PyTorch
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     @slow
     def test_lm_generate_ctrl(self):

tests/models/deformable_detr/test_modeling_deformable_detr.py

@@ -25,7 +25,7 @@ from transformers.file_utils import cached_property
 from transformers.testing_utils import (
     require_timm,
     require_torch,
-    require_torch_gpu,
+    require_torch_accelerator,
     require_vision,
     slow,
     torch_device,
@@ -648,7 +648,7 @@ class DeformableDetrModelIntegrationTests(unittest.TestCase):
         self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes)
         self.assertTrue(torch.allclose(outputs.pred_boxes[0, :3, :3], expected_boxes, atol=1e-4))
 
-    @require_torch_gpu
+    @require_torch_accelerator
     def test_inference_object_detection_head_equivalence_cpu_gpu(self):
         image_processor = self.default_image_processor
         image = prepare_img()
@@ -663,10 +663,10 @@ class DeformableDetrModelIntegrationTests(unittest.TestCase):
             cpu_outputs = model(pixel_values, pixel_mask)
 
         # 2. run model on GPU
-        model.to("cuda")
+        model.to(torch_device)
 
         with torch.no_grad():
-            gpu_outputs = model(pixel_values.to("cuda"), pixel_mask.to("cuda"))
+            gpu_outputs = model(pixel_values.to(torch_device), pixel_mask.to(torch_device))
 
         # 3. assert equivalence
         for key in cpu_outputs.keys():

tests/models/deit/test_modeling_deit.py

@@ -24,7 +24,8 @@ from transformers.models.auto import get_values
 from transformers.testing_utils import (
     require_accelerate,
     require_torch,
-    require_torch_gpu,
+    require_torch_accelerator,
+    require_torch_fp16,
     require_vision,
     slow,
     torch_device,
@@ -424,7 +425,8 @@ class DeiTModelIntegrationTest(unittest.TestCase):
 
     @slow
     @require_accelerate
-    @require_torch_gpu
+    @require_torch_accelerator
+    @require_torch_fp16
     def test_inference_fp16(self):
         r"""
         A small test to make sure that inference work in half precision without any problem.

tests/models/distilbert/test_modeling_distilbert.py

@@ -17,7 +17,7 @@ import tempfile
 import unittest
 
 from transformers import DistilBertConfig, is_torch_available
-from transformers.testing_utils import require_torch, require_torch_gpu, slow, torch_device
+from transformers.testing_utils import require_torch, require_torch_accelerator, slow, torch_device
 
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
@@ -264,7 +264,7 @@ class DistilBertModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCa
             self.assertIsNotNone(model)
 
     @slow
-    @require_torch_gpu
+    @require_torch_accelerator
     def test_torchscript_device_change(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
         for model_class in self.all_model_classes:

tests/models/ernie/test_modeling_ernie.py

@@ -18,7 +18,7 @@ import unittest
 
 from transformers import ErnieConfig, is_torch_available
 from transformers.models.auto import get_values
-from transformers.testing_utils import require_torch, require_torch_gpu, slow, torch_device
+from transformers.testing_utils import require_torch, require_torch_accelerator, slow, torch_device
 
 from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
@@ -574,7 +574,7 @@ class ErnieModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
             self.assertIsNotNone(model)
 
     @slow
-    @require_torch_gpu
+    @require_torch_accelerator
     def test_torchscript_device_change(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
         for model_class in self.all_model_classes:

tests/models/flaubert/test_modeling_flaubert.py

@@ -17,7 +17,7 @@ import tempfile
 import unittest
 
 from transformers import FlaubertConfig, is_torch_available
-from transformers.testing_utils import require_torch, require_torch_gpu, slow, torch_device
+from transformers.testing_utils import require_torch, require_torch_accelerator, slow, torch_device
 
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
@@ -463,7 +463,7 @@ class FlaubertModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase
             self.assertIsNotNone(model)
 
     @slow
-    @require_torch_gpu
+    @require_torch_accelerator
     def test_torchscript_device_change(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
         for model_class in self.all_model_classes:

tests/models/fsmt/test_modeling_fsmt.py

@@ -20,7 +20,14 @@ import timeout_decorator  # noqa
 from parameterized import parameterized
 
 from transformers import FSMTConfig, 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
@@ -398,11 +405,11 @@ class FSMTHeadTests(unittest.TestCase):
         self.assertEqual(n_pad_after, n_pad_before - 1)
         self.assertTrue(torch.eq(shifted[:, 0], 2).all())
 
+    @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 = FSMTForConditionalGeneration(config).eval().to(torch_device)
-        if torch_device == "cuda":
         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)
@@ -538,8 +545,7 @@ class FSMTModelIntegrationTests(unittest.TestCase):
     @slow
     def test_translation_pipeline(self, pair):
         tokenizer, model, src_text, tgt_text = self.translation_setup(pair)
-        device = 0 if torch_device == "cuda" else -1
-        pipeline = TranslationPipeline(model, tokenizer, framework="pt", device=device)
+        pipeline = TranslationPipeline(model, tokenizer, framework="pt", device=torch_device)
         output = pipeline([src_text])
         self.assertEqual([tgt_text], [x["translation_text"] for x in output])
 

tests/models/fuyu/test_modeling_fuyu.py

@@ -4,7 +4,7 @@ import unittest
 import requests
 
 from transformers import AutoTokenizer, FuyuConfig, is_torch_available, is_vision_available
-from transformers.testing_utils import require_torch, require_torch_gpu, slow, torch_device
+from transformers.testing_utils import require_torch, require_torch_accelerator, slow, torch_device
 
 from ...test_modeling_common import ids_tensor, random_attention_mask
 
@@ -257,7 +257,7 @@ class FuyuModelTester:
 
 
 @require_torch
-@require_torch_gpu
+@require_torch_accelerator
 @slow
 class FuyuIntegrationTest(unittest.TestCase):  # , ModelTesterMixin)
     """
@@ -279,7 +279,6 @@ class FuyuIntegrationTest(unittest.TestCase):  # , ModelTesterMixin)
         self.bus_image_pil = Image.open(io.BytesIO(requests.get(self.bus_image_url).content))
 
     @slow
-    @require_torch_gpu
     def test_model_8b_chat_greedy_generation_bus_captioning(self):
         EXPECTED_TEXT_COMPLETION = """A bus parked on the side of a road.|ENDOFTEXT|"""
         text_prompt_coco_captioning = "Generate a coco-style caption.\n"

tests/models/gpt2/test_modeling_gpt2.py

@@ -20,7 +20,7 @@ import math
 import unittest
 
 from transformers import GPT2Config, is_torch_available
-from transformers.testing_utils import require_torch, slow, torch_device
+from transformers.testing_utils import backend_empty_cache, require_torch, slow, torch_device
 
 from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
@@ -505,7 +505,7 @@ class GPT2ModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin
         super().tearDown()
         # clean-up as much as possible GPU memory occupied by PyTorch
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def test_config(self):
         self.config_tester.run_common_tests()
@@ -712,7 +712,7 @@ class GPT2ModelLanguageGenerationTest(unittest.TestCase):
         super().tearDown()
         # clean-up as much as possible GPU memory occupied by PyTorch
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def _test_lm_generate_gpt2_helper(
         self,

tests/models/gptj/test_modeling_gptj.py

@@ -562,7 +562,8 @@ class GPTJModelLanguageGenerationTest(unittest.TestCase):
         output_seq_strs = tokenizer.batch_decode(output_seq, skip_special_tokens=True)
         output_seq_tt_strs = tokenizer.batch_decode(output_seq_tt, skip_special_tokens=True)
 
-        if torch_device == "cuda":
+        if torch_device != "cpu":
+            # currently this expect value is only for `cuda`
             EXPECTED_OUTPUT_STR = (
                 "Today is a nice day and I've already been enjoying it. I walked to work with my wife"
             )