enable torchao test cases on XPU and switch to device agnostic APIs for test cases

src/diffusers/quantizers/quantization_config.py

@@ -656,13 +656,13 @@ def generate_fpx_quantization_types(bits: int):
 
     @staticmethod
     def _is_cuda_capability_atleast_8_9() -> bool:
-        if not torch.cuda.is_available():
-            raise RuntimeError("TorchAO requires a CUDA compatible GPU and installation of PyTorch.")
-
-        major, minor = torch.cuda.get_device_capability()
-        if major == 8:
-            return minor >= 9
-        return major >= 9
+        if torch.cuda.is_available():
+            major, minor = torch.cuda.get_device_capability()
+            if major == 8:
+                return minor >= 9
+            return major >= 9
+        else:
+            return True
 
     def get_apply_tensor_subclass(self):
         TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method()

src/diffusers/utils/testing_utils.py

@@ -300,9 +300,7 @@ def require_torch_gpu(test_case):
 
 def require_torch_cuda_compatibility(expected_compute_capability):
     def decorator(test_case):
-        if not torch.cuda.is_available():
-            return unittest.skip(test_case)
-        else:
+        if torch.cuda.is_available():
             current_compute_capability = get_torch_cuda_device_capability()
             return unittest.skipUnless(
                 float(current_compute_capability) == float(expected_compute_capability),

tests/models/autoencoders/test_models_consistency_decoder_vae.py

@@ -21,6 +21,7 @@
 
 from diffusers import ConsistencyDecoderVAE, StableDiffusionPipeline
 from diffusers.utils.testing_utils import (
+    backend_empty_cache,
     enable_full_determinism,
     load_image,
     slow,
@@ -162,13 +163,13 @@ def setUp(self):
         # clean up the VRAM before each test
         super().setUp()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def tearDown(self):
         # clean up the VRAM after each test
         super().tearDown()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     @torch.no_grad()
     def test_encode_decode(self):

tests/models/unets/test_models_unet_2d.py

@@ -22,6 +22,7 @@
 from diffusers import UNet2DModel
 from diffusers.utils import logging
 from diffusers.utils.testing_utils import (
+    backend_empty_cache,
     enable_full_determinism,
     floats_tensor,
     require_torch_accelerator,
@@ -229,7 +230,7 @@ def test_from_pretrained_accelerate_wont_change_results(self):
 
         # two models don't need to stay in the device at the same time
         del model_accelerate
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
         gc.collect()
 
         model_normal_load, _ = UNet2DModel.from_pretrained(

tests/models/unets/test_models_unet_2d_condition.py

@@ -46,7 +46,6 @@
     require_peft_backend,
     require_torch_accelerator,
     require_torch_accelerator_with_fp16,
-    require_torch_gpu,
     skip_mps,
     slow,
     torch_all_close,
@@ -980,13 +979,13 @@ def test_ip_adapter_plus(self):
         assert sample2.allclose(sample5, atol=1e-4, rtol=1e-4)
         assert sample2.allclose(sample6, atol=1e-4, rtol=1e-4)
 
-    @require_torch_gpu
     @parameterized.expand(
         [
             ("hf-internal-testing/unet2d-sharded-dummy", None),
             ("hf-internal-testing/tiny-sd-unet-sharded-latest-format", "fp16"),
         ]
     )
+    @require_torch_accelerator
     def test_load_sharded_checkpoint_from_hub(self, repo_id, variant):
         _, inputs_dict = self.prepare_init_args_and_inputs_for_common()
         loaded_model = self.model_class.from_pretrained(repo_id, variant=variant)
@@ -996,13 +995,13 @@ def test_load_sharded_checkpoint_from_hub(self, repo_id, variant):
         assert loaded_model
         assert new_output.sample.shape == (4, 4, 16, 16)
 
-    @require_torch_gpu
     @parameterized.expand(
         [
             ("hf-internal-testing/unet2d-sharded-dummy-subfolder", None),
             ("hf-internal-testing/tiny-sd-unet-sharded-latest-format-subfolder", "fp16"),
         ]
     )
+    @require_torch_accelerator
     def test_load_sharded_checkpoint_from_hub_subfolder(self, repo_id, variant):
         _, inputs_dict = self.prepare_init_args_and_inputs_for_common()
         loaded_model = self.model_class.from_pretrained(repo_id, subfolder="unet", variant=variant)

tests/pipelines/allegro/test_allegro.py

@@ -24,6 +24,7 @@
 
 from diffusers import AllegroPipeline, AllegroTransformer3DModel, AutoencoderKLAllegro, DDIMScheduler
 from diffusers.utils.testing_utils import (
+    backend_empty_cache,
     enable_full_determinism,
     numpy_cosine_similarity_distance,
     require_hf_hub_version_greater,
@@ -341,12 +342,12 @@ class AllegroPipelineIntegrationTests(unittest.TestCase):
     def setUp(self):
         super().setUp()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def test_allegro(self):
         generator = torch.Generator("cpu").manual_seed(0)

tests/pipelines/audioldm/test_audioldm.py

@@ -37,7 +37,7 @@
     UNet2DConditionModel,
 )
 from diffusers.utils import is_xformers_available
-from diffusers.utils.testing_utils import enable_full_determinism, nightly, torch_device
+from diffusers.utils.testing_utils import backend_empty_cache, enable_full_determinism, nightly, torch_device
 
 from ..pipeline_params import TEXT_TO_AUDIO_BATCH_PARAMS, TEXT_TO_AUDIO_PARAMS
 from ..test_pipelines_common import PipelineTesterMixin
@@ -378,12 +378,12 @@ class AudioLDMPipelineSlowTests(unittest.TestCase):
     def setUp(self):
         super().setUp()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
         generator = torch.Generator(device=generator_device).manual_seed(seed)
@@ -423,12 +423,12 @@ class AudioLDMPipelineNightlyTests(unittest.TestCase):
     def setUp(self):
         super().setUp()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
         generator = torch.Generator(device=generator_device).manual_seed(seed)

tests/pipelines/audioldm2/test_audioldm2.py

@@ -45,7 +45,13 @@
     LMSDiscreteScheduler,
     PNDMScheduler,
 )
-from diffusers.utils.testing_utils import enable_full_determinism, is_torch_version, nightly, torch_device
+from diffusers.utils.testing_utils import (
+    backend_empty_cache,
+    enable_full_determinism,
+    is_torch_version,
+    nightly,
+    torch_device,
+)
 
 from ..pipeline_params import TEXT_TO_AUDIO_BATCH_PARAMS, TEXT_TO_AUDIO_PARAMS
 from ..test_pipelines_common import PipelineTesterMixin
@@ -540,12 +546,12 @@ class AudioLDM2PipelineSlowTests(unittest.TestCase):
     def setUp(self):
         super().setUp()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
         generator = torch.Generator(device=generator_device).manual_seed(seed)

tests/pipelines/cogvideo/test_cogvideox.py

@@ -22,6 +22,7 @@
 
 from diffusers import AutoencoderKLCogVideoX, CogVideoXPipeline, CogVideoXTransformer3DModel, DDIMScheduler
 from diffusers.utils.testing_utils import (
+    backend_empty_cache,
     enable_full_determinism,
     numpy_cosine_similarity_distance,
     require_torch_accelerator,
@@ -334,12 +335,12 @@ class CogVideoXPipelineIntegrationTests(unittest.TestCase):
     def setUp(self):
         super().setUp()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def test_cogvideox(self):
         generator = torch.Generator("cpu").manual_seed(0)

tests/pipelines/cogview3/test_cogview3plus.py

@@ -22,6 +22,7 @@
 
 from diffusers import AutoencoderKL, CogVideoXDDIMScheduler, CogView3PlusPipeline, CogView3PlusTransformer2DModel
 from diffusers.utils.testing_utils import (
+    backend_empty_cache,
     enable_full_determinism,
     numpy_cosine_similarity_distance,
     require_torch_accelerator,
@@ -244,12 +245,12 @@ class CogView3PlusPipelineIntegrationTests(unittest.TestCase):
     def setUp(self):
         super().setUp()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def test_cogview3plus(self):
         generator = torch.Generator("cpu").manual_seed(0)

tests/pipelines/controlnet/test_controlnet_img2img.py

@@ -36,6 +36,7 @@
 from diffusers.utils import load_image
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.testing_utils import (
+    backend_empty_cache,
     enable_full_determinism,
     floats_tensor,
     load_numpy,
@@ -412,12 +413,12 @@ class ControlNetImg2ImgPipelineSlowTests(unittest.TestCase):
     def setUp(self):
         super().setUp()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def test_canny(self):
         controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny")

tests/pipelines/controlnet/test_controlnet_inpaint.py

@@ -36,6 +36,7 @@
 from diffusers.utils import load_image
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.testing_utils import (
+    backend_empty_cache,
     enable_full_determinism,
     floats_tensor,
     load_numpy,
@@ -464,12 +465,12 @@ class ControlNetInpaintPipelineSlowTests(unittest.TestCase):
     def setUp(self):
         super().setUp()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def test_canny(self):
         controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny")

tests/pipelines/controlnet_sd3/test_controlnet_sd3.py

@@ -221,7 +221,7 @@ def test_xformers_attention_forwardGenerator_pass(self):
 
 @slow
 @require_big_accelerator
-@pytest.mark.big_gpu_with_torch_cuda
+@pytest.mark.big_accelerator
 class StableDiffusion3ControlNetPipelineSlowTests(unittest.TestCase):
     pipeline_class = StableDiffusion3ControlNetPipeline
 

tests/pipelines/deepfloyd_if/test_if.py

@@ -25,6 +25,7 @@
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.testing_utils import (
     backend_empty_cache,
+    backend_max_memory_allocated,
     backend_reset_max_memory_allocated,
     backend_reset_peak_memory_stats,
     load_numpy,
@@ -135,7 +136,7 @@ def test_if_text_to_image(self):
 
         image = output.images[0]
 
-        mem_bytes = torch.cuda.max_memory_allocated()
+        mem_bytes = backend_max_memory_allocated(torch_device)
         assert mem_bytes < 12 * 10**9
 
         expected_image = load_numpy(

tests/pipelines/deepfloyd_if/test_if_img2img.py

@@ -24,6 +24,7 @@
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.testing_utils import (
     backend_empty_cache,
+    backend_max_memory_allocated,
     backend_reset_max_memory_allocated,
     backend_reset_peak_memory_stats,
     floats_tensor,
@@ -151,7 +152,7 @@ def test_if_img2img(self):
         )
         image = output.images[0]
 
-        mem_bytes = torch.cuda.max_memory_allocated()
+        mem_bytes = backend_max_memory_allocated(torch_device)
         assert mem_bytes < 12 * 10**9
 
         expected_image = load_numpy(

tests/pipelines/flux/test_pipeline_flux.py

@@ -224,7 +224,7 @@ def test_flux_true_cfg(self):
 
 @nightly
 @require_big_accelerator
-@pytest.mark.big_gpu_with_torch_cuda
+@pytest.mark.big_accelerator
 class FluxPipelineSlowTests(unittest.TestCase):
     pipeline_class = FluxPipeline
     repo_id = "black-forest-labs/FLUX.1-schnell"
@@ -312,7 +312,7 @@ def test_flux_inference(self):
 
 @slow
 @require_big_accelerator
-@pytest.mark.big_gpu_with_torch_cuda
+@pytest.mark.big_accelerator
 class FluxIPAdapterPipelineSlowTests(unittest.TestCase):
     pipeline_class = FluxPipeline
     repo_id = "black-forest-labs/FLUX.1-dev"

tests/pipelines/flux/test_pipeline_flux_redux.py

@@ -19,7 +19,7 @@
 
 @slow
 @require_big_accelerator
-@pytest.mark.big_gpu_with_torch_cuda
+@pytest.mark.big_accelerator
 class FluxReduxSlowTests(unittest.TestCase):
     pipeline_class = FluxPriorReduxPipeline
     repo_id = "black-forest-labs/FLUX.1-Redux-dev"

tests/pipelines/hunyuandit/test_hunyuan_dit.py

@@ -23,6 +23,7 @@
 
 from diffusers import AutoencoderKL, DDPMScheduler, HunyuanDiT2DModel, HunyuanDiTPipeline
 from diffusers.utils.testing_utils import (
+    backend_empty_cache,
     enable_full_determinism,
     numpy_cosine_similarity_distance,
     require_torch_accelerator,
@@ -310,12 +311,12 @@ class HunyuanDiTPipelineIntegrationTests(unittest.TestCase):
     def setUp(self):
         super().setUp()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        torch.cuda.empty_cache()
+        backend_empty_cache(torch_device)
 
     def test_hunyuan_dit_1024(self):
         generator = torch.Generator("cpu").manual_seed(0)