make pipelines tests device-agnostic (part1)

src/diffusers/utils/testing_utils.py

@@ -373,6 +373,14 @@ def require_note_seq(test_case):
     return unittest.skipUnless(is_note_seq_available(), "test requires note_seq")(test_case)
 
 
+def require_non_cpu(test_case):
+    """
+    Decorator marking a test that requires a hardware accelerator backend. These tests are skipped when there are no
+    hardware accelerator available.
+    """
+    return unittest.skipUnless(torch_device != "cpu", "test requires a hardware accelerator")(test_case)
+
+
 def require_torchsde(test_case):
     """
     Decorator marking a test that requires torchsde. These tests are skipped when torchsde isn't installed.

tests/pipelines/amused/test_amused.py

@@ -22,7 +22,7 @@
 from diffusers import AmusedPipeline, AmusedScheduler, UVit2DModel, VQModel
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
-    require_torch_gpu,
+    require_torch_accelerator,
     slow,
     torch_device,
 )
@@ -129,7 +129,7 @@ def test_inference_batch_single_identical(self):
 
 
 @slow
-@require_torch_gpu
+@require_torch_accelerator
 class AmusedPipelineSlowTests(unittest.TestCase):
     def test_amused_256(self):
         pipe = AmusedPipeline.from_pretrained("amused/amused-256")

tests/pipelines/amused/test_amused_img2img.py

@@ -23,7 +23,7 @@
 from diffusers.utils import load_image
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
-    require_torch_gpu,
+    require_torch_accelerator,
     slow,
     torch_device,
 )
@@ -131,7 +131,7 @@ def test_inference_batch_single_identical(self):
 
 
 @slow
-@require_torch_gpu
+@require_torch_accelerator
 class AmusedImg2ImgPipelineSlowTests(unittest.TestCase):
     def test_amused_256(self):
         pipe = AmusedImg2ImgPipeline.from_pretrained("amused/amused-256")

tests/pipelines/amused/test_amused_inpaint.py

@@ -23,7 +23,7 @@
 from diffusers.utils import load_image
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
-    require_torch_gpu,
+    require_torch_accelerator,
     slow,
     torch_device,
 )
@@ -135,7 +135,7 @@ def test_inference_batch_single_identical(self):
 
 
 @slow
-@require_torch_gpu
+@require_torch_accelerator
 class AmusedInpaintPipelineSlowTests(unittest.TestCase):
     def test_amused_256(self):
         pipe = AmusedInpaintPipeline.from_pretrained("amused/amused-256")

tests/pipelines/animatediff/test_animatediff.py

@@ -19,7 +19,13 @@
 )
 from diffusers.models.attention import FreeNoiseTransformerBlock
 from diffusers.utils import is_xformers_available, logging
-from diffusers.utils.testing_utils import numpy_cosine_similarity_distance, require_torch_gpu, slow, torch_device
+from diffusers.utils.testing_utils import (
+    numpy_cosine_similarity_distance,
+    require_non_cpu,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
 from ..test_pipelines_common import (
@@ -272,7 +278,7 @@ def test_inference_batch_single_identical(
         max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
         assert max_diff < expected_max_diff
 
-    @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
+    @require_non_cpu
     def test_to_device(self):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
@@ -288,14 +294,14 @@ def test_to_device(self):
         output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
         self.assertTrue(np.isnan(output_cpu).sum() == 0)
 
-        pipe.to("cuda")
+        pipe.to(torch_device)
         model_devices = [
             component.device.type for component in pipe.components.values() if hasattr(component, "device")
         ]
-        self.assertTrue(all(device == "cuda" for device in model_devices))
+        self.assertTrue(all(device == torch_device for device in model_devices))
 
-        output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
-        self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)
+        output_device = pipe(**self.get_dummy_inputs(torch_device))[0]
+        self.assertTrue(np.isnan(to_np(output_device)).sum() == 0)
 
     def test_to_dtype(self):
         components = self.get_dummy_components()

tests/pipelines/animatediff/test_animatediff_controlnet.py

@@ -21,7 +21,7 @@
 from diffusers.models.attention import FreeNoiseTransformerBlock
 from diffusers.utils import logging
 from diffusers.utils.import_utils import is_xformers_available
-from diffusers.utils.testing_utils import torch_device
+from diffusers.utils.testing_utils import require_non_cpu, torch_device
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
 from ..test_pipelines_common import (
@@ -281,7 +281,7 @@ def test_inference_batch_single_identical(
         max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
         assert max_diff < expected_max_diff
 
-    @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
+    @require_non_cpu
     def test_to_device(self):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
@@ -297,14 +297,14 @@ def test_to_device(self):
         output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
         self.assertTrue(np.isnan(output_cpu).sum() == 0)
 
-        pipe.to("cuda")
+        pipe.to(torch_device)
         model_devices = [
             component.device.type for component in pipe.components.values() if hasattr(component, "device")
         ]
-        self.assertTrue(all(device == "cuda" for device in model_devices))
+        self.assertTrue(all(device == torch_device for device in model_devices))
 
-        output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
-        self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)
+        output_device = pipe(**self.get_dummy_inputs(torch_device))[0]
+        self.assertTrue(np.isnan(to_np(output_device)).sum() == 0)
 
     def test_to_dtype(self):
         components = self.get_dummy_components()

tests/pipelines/animatediff/test_animatediff_sdxl.py

@@ -14,7 +14,7 @@
     UNetMotionModel,
 )
 from diffusers.utils import is_xformers_available, logging
-from diffusers.utils.testing_utils import torch_device
+from diffusers.utils.testing_utils import require_non_cpu, torch_device
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS, TEXT_TO_IMAGE_PARAMS
 from ..test_pipelines_common import (
@@ -212,7 +212,7 @@ def test_inference_batch_single_identical(
         max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
         assert max_diff < expected_max_diff
 
-    @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
+    @require_non_cpu
     def test_to_device(self):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
@@ -228,14 +228,14 @@ def test_to_device(self):
         output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
         self.assertTrue(np.isnan(output_cpu).sum() == 0)
 
-        pipe.to("cuda")
+        pipe.to(torch_device)
         model_devices = [
             component.device.type for component in pipe.components.values() if hasattr(component, "device")
         ]
-        self.assertTrue(all(device == "cuda" for device in model_devices))
+        self.assertTrue(all(device == torch_device for device in model_devices))
 
-        output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
-        self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)
+        output_device = pipe(**self.get_dummy_inputs(torch_device))[0]
+        self.assertTrue(np.isnan(to_np(output_device)).sum() == 0)
 
     def test_to_dtype(self):
         components = self.get_dummy_components()

tests/pipelines/animatediff/test_animatediff_sparsectrl.py

@@ -20,7 +20,7 @@
 )
 from diffusers.utils import logging
 from diffusers.utils.import_utils import is_xformers_available
-from diffusers.utils.testing_utils import torch_device
+from diffusers.utils.testing_utils import require_non_cpu, torch_device
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
 from ..test_pipelines_common import (
@@ -345,7 +345,7 @@ def test_inference_batch_single_identical_use_simplified_condition_embedding_tru
         max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
         assert max_diff < expected_max_diff
 
-    @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
+    @require_non_cpu
     def test_to_device(self):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
@@ -361,13 +361,13 @@ def test_to_device(self):
         output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
         self.assertTrue(np.isnan(output_cpu).sum() == 0)
 
-        pipe.to("cuda")
+        pipe.to(torch_device)
         model_devices = [
             component.device.type for component in pipe.components.values() if hasattr(component, "device")
         ]
-        self.assertTrue(all(device == "cuda" for device in model_devices))
+        self.assertTrue(all(device == torch_device for device in model_devices))
 
-        output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
+        output_cuda = pipe(**self.get_dummy_inputs(torch_device))[0]
         self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)
 
     def test_to_dtype(self):

tests/pipelines/animatediff/test_animatediff_video2video.py

@@ -19,7 +19,7 @@
 )
 from diffusers.models.attention import FreeNoiseTransformerBlock
 from diffusers.utils import is_xformers_available, logging
-from diffusers.utils.testing_utils import torch_device
+from diffusers.utils.testing_utils import require_non_cpu, torch_device
 
 from ..pipeline_params import TEXT_TO_IMAGE_PARAMS, VIDEO_TO_VIDEO_BATCH_PARAMS
 from ..test_pipelines_common import IPAdapterTesterMixin, PipelineFromPipeTesterMixin, PipelineTesterMixin
@@ -258,7 +258,7 @@ def test_inference_batch_single_identical(
         max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
         assert max_diff < expected_max_diff
 
-    @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
+    @require_non_cpu
     def test_to_device(self):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
@@ -274,14 +274,14 @@ def test_to_device(self):
         output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
         self.assertTrue(np.isnan(output_cpu).sum() == 0)
 
-        pipe.to("cuda")
+        pipe.to(torch_device)
         model_devices = [
             component.device.type for component in pipe.components.values() if hasattr(component, "device")
         ]
-        self.assertTrue(all(device == "cuda" for device in model_devices))
+        self.assertTrue(all(device == torch_device for device in model_devices))
 
-        output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
-        self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)
+        output_device = pipe(**self.get_dummy_inputs(torch_device))[0]
+        self.assertTrue(np.isnan(to_np(output_device)).sum() == 0)
 
     def test_to_dtype(self):
         components = self.get_dummy_components()

tests/pipelines/animatediff/test_animatediff_video2video_controlnet.py

@@ -20,7 +20,7 @@
 )
 from diffusers.models.attention import FreeNoiseTransformerBlock
 from diffusers.utils import is_xformers_available, logging
-from diffusers.utils.testing_utils import torch_device
+from diffusers.utils.testing_utils import require_non_cpu, torch_device
 
 from ..pipeline_params import TEXT_TO_IMAGE_PARAMS, VIDEO_TO_VIDEO_BATCH_PARAMS
 from ..test_pipelines_common import IPAdapterTesterMixin, PipelineFromPipeTesterMixin, PipelineTesterMixin
@@ -274,7 +274,7 @@ def test_inference_batch_single_identical(
         max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
         assert max_diff < expected_max_diff
 
-    @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
+    @require_non_cpu
     def test_to_device(self):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
@@ -290,13 +290,13 @@ def test_to_device(self):
         output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
         self.assertTrue(np.isnan(output_cpu).sum() == 0)
 
-        pipe.to("cuda")
+        pipe.to(torch_device)
         model_devices = [
             component.device.type for component in pipe.components.values() if hasattr(component, "device")
         ]
-        self.assertTrue(all(device == "cuda" for device in model_devices))
+        self.assertTrue(all(device == torch_device for device in model_devices))
 
-        output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
+        output_cuda = pipe(**self.get_dummy_inputs(torch_device))[0]
         self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)
 
     def test_to_dtype(self):

tests/pipelines/controlnet_xs/test_controlnetxs.py

@@ -38,6 +38,7 @@
     is_torch_compile,
     load_image,
     load_numpy,
+    require_non_cpu,
     require_torch_2,
     require_torch_gpu,
     run_test_in_subprocess,
@@ -306,7 +307,7 @@ def test_multi_vae(self):
 
             assert out_vae_np.shape == out_np.shape
 
-    @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
+    @require_non_cpu
     def test_to_device(self):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
@@ -322,14 +323,14 @@ def test_to_device(self):
         output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
         self.assertTrue(np.isnan(output_cpu).sum() == 0)
 
-        pipe.to("cuda")
+        pipe.to(torch_device)
         model_devices = [
             component.device.type for component in pipe.components.values() if hasattr(component, "device")
         ]
-        self.assertTrue(all(device == "cuda" for device in model_devices))
+        self.assertTrue(all(device == torch_device for device in model_devices))
 
-        output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
-        self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)
+        output_device = pipe(**self.get_dummy_inputs(torch_device))[0]
+        self.assertTrue(np.isnan(to_np(output_device)).sum() == 0)
 
 
 @slow

tests/pipelines/deepfloyd_if/test_if.py

@@ -23,7 +23,7 @@
 )
 from diffusers.models.attention_processor import AttnAddedKVProcessor
 from diffusers.utils.import_utils import is_xformers_available
-from diffusers.utils.testing_utils import load_numpy, require_torch_gpu, skip_mps, slow, torch_device
+from diffusers.utils.testing_utils import load_numpy, require_non_cpu, require_torch_gpu, skip_mps, slow, torch_device
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
 from ..test_pipelines_common import PipelineTesterMixin, assert_mean_pixel_difference
@@ -58,7 +58,7 @@ def get_dummy_inputs(self, device, seed=0):
     def test_save_load_optional_components(self):
         self._test_save_load_optional_components()
 
-    @unittest.skipIf(torch_device != "cuda", reason="float16 requires CUDA")
+    @require_non_cpu
     def test_save_load_float16(self):
         # Due to non-determinism in save load of the hf-internal-testing/tiny-random-t5 text encoder
         super().test_save_load_float16(expected_max_diff=1e-1)

tests/pipelines/deepfloyd_if/test_if_img2img.py

@@ -22,7 +22,15 @@
 from diffusers import IFImg2ImgPipeline
 from diffusers.models.attention_processor import AttnAddedKVProcessor
 from diffusers.utils.import_utils import is_xformers_available
-from diffusers.utils.testing_utils import floats_tensor, load_numpy, require_torch_gpu, skip_mps, slow, torch_device
+from diffusers.utils.testing_utils import (
+    floats_tensor,
+    load_numpy,
+    require_non_cpu,
+    require_torch_gpu,
+    skip_mps,
+    slow,
+    torch_device,
+)
 
 from ..pipeline_params import (
     TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS,
@@ -70,12 +78,12 @@ def test_save_load_optional_components(self):
     def test_xformers_attention_forwardGenerator_pass(self):
         self._test_xformers_attention_forwardGenerator_pass(expected_max_diff=1e-3)
 
-    @unittest.skipIf(torch_device != "cuda", reason="float16 requires CUDA")
+    @require_non_cpu
     def test_save_load_float16(self):
         # Due to non-determinism in save load of the hf-internal-testing/tiny-random-t5 text encoder
         super().test_save_load_float16(expected_max_diff=1e-1)
 
-    @unittest.skipIf(torch_device != "cuda", reason="float16 requires CUDA")
+    @require_non_cpu
     def test_float16_inference(self):
         super().test_float16_inference(expected_max_diff=1e-1)