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 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 @@ def test_beam_search_generate_dict_outputs_use_cache(self):
                 )
 
     @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 @@
     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
@@ -570,13 +576,13 @@ def test_inputs_embeds(self):
             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.half()
         model.generate(input_ids, attention_mask=attention_mask)
         model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
 
@@ -636,13 +642,13 @@ def test_inputs_embeds(self):
             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.half()
         model.generate(input_ids, attention_mask=attention_mask)
         model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
 
@@ -700,14 +706,14 @@ def test_inputs_embeds(self):
             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()
+        model.half()
 
         # toy generation_configs
         semantic_generation_config = BarkSemanticGenerationConfig(semantic_vocab_size=0)

tests/models/bart/test_modeling_bart.py

@@ -22,7 +22,14 @@
 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
@@ -383,12 +390,12 @@ def test_tokenization(self):
             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.half()
         model.generate(input_ids, attention_mask=attention_mask)
         model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
 
@@ -497,13 +504,13 @@ def test_inputs_embeds(self):
             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.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 @@
 
 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 @@ def test_model_from_pretrained(self):
             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 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
@@ -376,13 +383,13 @@ def test_inputs_embeds(self):
             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.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 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
@@ -269,13 +277,13 @@ def test_encoder_decoder_model_standalone(self):
         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.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 @@ def test_generation_from_short_input_same_as_parlai_3B(self):
         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 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
@@ -269,13 +274,13 @@ def test_encoder_decoder_model_standalone(self):
         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.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 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 @@ def test_inference_image_captioning(self):
             [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 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 @@ def test_inference_t5_batched_beam_search(self):
         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 @@ def test_inference_opt_multi_gpu(self):
         )
         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 @@
 
 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 @@ def test_attention_outputs(self):
                 )
 
     @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 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 @@ def tearDown(self):
         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 @@ def tearDown(self):
         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.testing_utils import (
     require_timm,
     require_torch,
-    require_torch_gpu,
+    require_torch_accelerator,
     require_vision,
     slow,
     torch_device,
@@ -648,7 +648,7 @@ def test_inference_object_detection_head_with_box_refine_two_stage(self):
         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 @@ def test_inference_object_detection_head_equivalence_cpu_gpu(self):
             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():