[Tests] Speed up fast pipelines part II (#7521)

* start printing the tensors.

* print full throttle

* set static slices for 7 tests.

* remove printing.

* flatten

* disable test for controlnet

* what happens when things are seeded properly?

* set the right value

* style./

* make pia test fail to check things

* print.

* fix pia.

* checking for animatediff.

* fix: animatediff.

* video synthesis

* final piece.

* style.

* print guess.

* fix: assertion for control guess.

---------

Co-authored-by: Dhruv Nair <[email protected]>

Author: sayakpaul

tests/pipelines/animatediff/test_animatediff.py

@@ -172,6 +172,12 @@ def test_ip_adapter_single(self):
             )
         return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
 
+    def test_dict_tuple_outputs_equivalent(self):
+        expected_slice = None
+        if torch_device == "cpu":
+            expected_slice = np.array([0.4051, 0.4495, 0.4480, 0.5845, 0.4172, 0.6066, 0.4205, 0.3786, 0.5323])
+        return super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+
     def test_inference_batch_single_identical(
         self,
         batch_size=2,

tests/pipelines/controlnet/test_controlnet_blip_diffusion.py

@@ -28,7 +28,7 @@
     PNDMScheduler,
     UNet2DConditionModel,
 )
-from diffusers.utils.testing_utils import enable_full_determinism
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
 from src.diffusers.pipelines.blip_diffusion.blip_image_processing import BlipImageProcessor
 from src.diffusers.pipelines.blip_diffusion.modeling_blip2 import Blip2QFormerModel
 from src.diffusers.pipelines.blip_diffusion.modeling_ctx_clip import ContextCLIPTextModel
@@ -196,6 +196,12 @@ def get_dummy_inputs(self, device, seed=0):
         }
         return inputs
 
+    def test_dict_tuple_outputs_equivalent(self):
+        expected_slice = None
+        if torch_device == "cpu":
+            expected_slice = np.array([0.4803, 0.3865, 0.1422, 0.6119, 0.2283, 0.6365, 0.5453, 0.5205, 0.3581])
+        super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+
     def test_blipdiffusion_controlnet(self):
         device = "cpu"
         components = self.get_dummy_components()

tests/pipelines/controlnet/test_controlnet_inpaint_sdxl.py

@@ -57,6 +57,7 @@ class ControlNetPipelineSDXLFastTests(
     image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
 
     def get_dummy_components(self):
+        torch.manual_seed(0)
         unet = UNet2DConditionModel(
             block_out_channels=(32, 64),
             layers_per_block=2,
@@ -74,6 +75,7 @@ def get_dummy_components(self):
             projection_class_embeddings_input_dim=80,  # 6 * 8 + 32
             cross_attention_dim=64,
         )
+        torch.manual_seed(0)
         controlnet = ControlNetModel(
             block_out_channels=(32, 64),
             layers_per_block=2,
@@ -123,6 +125,7 @@ def get_dummy_components(self):
         text_encoder = CLIPTextModel(text_encoder_config)
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
 
+        torch.manual_seed(0)
         text_encoder_2 = CLIPTextModelWithProjection(text_encoder_config)
         tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
 
@@ -182,6 +185,12 @@ def get_dummy_inputs(self, device, seed=0, img_res=64):
     def test_attention_slicing_forward_pass(self):
         return self._test_attention_slicing_forward_pass(expected_max_diff=2e-3)
 
+    def test_dict_tuple_outputs_equivalent(self):
+        expected_slice = None
+        if torch_device == "cpu":
+            expected_slice = np.array([0.5490, 0.5053, 0.4676, 0.5816, 0.5364, 0.4830, 0.5937, 0.5719, 0.4318])
+        super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+
     @unittest.skipIf(
         torch_device != "cuda" or not is_xformers_available(),
         reason="XFormers attention is only available with CUDA and `xformers` installed",
@@ -289,9 +298,7 @@ def test_controlnet_sdxl_guess(self):
 
         output = sd_pipe(**inputs)
         image_slice = output.images[0, -3:, -3:, -1]
-        expected_slice = np.array(
-            [0.5381963, 0.4836803, 0.45821992, 0.5577731, 0.51210403, 0.4794795, 0.59282357, 0.5647199, 0.43100584]
-        )
+        expected_slice = np.array([0.549, 0.5053, 0.4676, 0.5816, 0.5364, 0.483, 0.5937, 0.5719, 0.4318])
 
         # make sure that it's equal
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-4

tests/pipelines/kandinsky3/test_kandinsky3_img2img.py

@@ -36,6 +36,7 @@
     load_image,
     require_torch_gpu,
     slow,
+    torch_device,
 )
 
 from ..pipeline_params import (
@@ -152,6 +153,12 @@ def get_dummy_inputs(self, device, seed=0):
         }
         return inputs
 
+    def test_dict_tuple_outputs_equivalent(self):
+        expected_slice = None
+        if torch_device == "cpu":
+            expected_slice = np.array([0.5762, 0.6112, 0.4150, 0.6018, 0.6167, 0.4626, 0.5426, 0.5641, 0.6536])
+        super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+
     def test_kandinsky3_img2img(self):
         device = "cpu"
 

tests/pipelines/pia/test_pia.py

@@ -175,6 +175,12 @@ def test_ip_adapter_single(self):
             )
         return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
 
+    def test_dict_tuple_outputs_equivalent(self):
+        expected_slice = None
+        if torch_device == "cpu":
+            expected_slice = np.array([0.3740, 0.4284, 0.4038, 0.5417, 0.4405, 0.5521, 0.4273, 0.4124, 0.4997])
+        return super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+
     @unittest.skip("Attention slicing is not enabled in this pipeline")
     def test_attention_slicing_forward_pass(self):
         pass

tests/pipelines/stable_diffusion_2/test_stable_diffusion_attend_and_excite.py

@@ -32,6 +32,7 @@
     numpy_cosine_similarity_distance,
     require_torch_gpu,
     skip_mps,
+    torch_device,
 )
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
@@ -153,6 +154,12 @@ def get_dummy_inputs(self, device, seed=0):
         }
         return inputs
 
+    def test_dict_tuple_outputs_equivalent(self):
+        expected_slice = None
+        if torch_device == "cpu":
+            expected_slice = np.array([0.6391, 0.6290, 0.4860, 0.5134, 0.5550, 0.4577, 0.5033, 0.5023, 0.4538])
+        super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice, expected_max_difference=3e-3)
+
     def test_inference(self):
         device = "cpu"
 
@@ -182,9 +189,6 @@ def test_inference_batch_consistent(self):
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(batch_size=2, expected_max_diff=7e-4)
 
-    def test_dict_tuple_outputs_equivalent(self):
-        super().test_dict_tuple_outputs_equivalent(expected_max_difference=3e-3)
-
     def test_pt_np_pil_outputs_equivalent(self):
         super().test_pt_np_pil_outputs_equivalent(expected_max_diff=5e-4)
 

tests/pipelines/stable_diffusion_gligen_text_image/test_stable_diffusion_gligen_text_image.py

@@ -35,7 +35,7 @@
 )
 from diffusers.pipelines.stable_diffusion import CLIPImageProjection
 from diffusers.utils import load_image
-from diffusers.utils.testing_utils import enable_full_determinism
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
 
 from ..pipeline_params import (
     TEXT_TO_IMAGE_BATCH_PARAMS,
@@ -160,6 +160,12 @@ def get_dummy_inputs(self, device, seed=0):
         }
         return inputs
 
+    def test_dict_tuple_outputs_equivalent(self):
+        expected_slice = None
+        if torch_device == "cpu":
+            expected_slice = np.array([0.5052, 0.5546, 0.4567, 0.4770, 0.5195, 0.4085, 0.5026, 0.4909, 0.4495])
+        super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+
     def test_stable_diffusion_gligen_text_image_default_case(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()

tests/pipelines/test_pipelines_common.py

@@ -1037,7 +1037,7 @@ 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
 
-    def test_dict_tuple_outputs_equivalent(self, expected_max_difference=1e-4):
+    def test_dict_tuple_outputs_equivalent(self, expected_slice=None, expected_max_difference=1e-4):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
         for component in pipe.components.values():
@@ -1048,10 +1048,21 @@ def test_dict_tuple_outputs_equivalent(self, expected_max_difference=1e-4):
         pipe.set_progress_bar_config(disable=None)
 
         generator_device = "cpu"
-        output = pipe(**self.get_dummy_inputs(generator_device))[0]
+        if expected_slice is None:
+            output = pipe(**self.get_dummy_inputs(generator_device))[0]
+        else:
+            output = expected_slice
+
         output_tuple = pipe(**self.get_dummy_inputs(generator_device), return_dict=False)[0]
 
-        max_diff = np.abs(to_np(output) - to_np(output_tuple)).max()
+        if expected_slice is None:
+            max_diff = np.abs(to_np(output) - to_np(output_tuple)).max()
+        else:
+            if output_tuple.ndim != 5:
+                max_diff = np.abs(to_np(output) - to_np(output_tuple)[0, -3:, -3:, -1].flatten()).max()
+            else:
+                max_diff = np.abs(to_np(output) - to_np(output_tuple)[0, -3:, -3:, -1, -1].flatten()).max()
+
         self.assertLess(max_diff, expected_max_difference)
 
     def test_components_function(self):

tests/pipelines/text_to_video_synthesis/test_text_to_video.py

@@ -133,6 +133,12 @@ def get_dummy_inputs(self, device, seed=0):
         }
         return inputs
 
+    def test_dict_tuple_outputs_equivalent(self):
+        expected_slice = None
+        if torch_device == "cpu":
+            expected_slice = np.array([0.4903, 0.5649, 0.5504, 0.5179, 0.4821, 0.5466, 0.4131, 0.5052, 0.5077])
+        return super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+
     def test_text_to_video_default_case(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()

tests/pipelines/unidiffuser/test_unidiffuser.py

@@ -206,6 +206,12 @@ def get_dummy_inputs_with_latents(self, device, seed=0):
         }
         return inputs
 
+    def test_dict_tuple_outputs_equivalent(self):
+        expected_slice = None
+        if torch_device == "cpu":
+            expected_slice = np.array([0.7489, 0.3722, 0.4475, 0.5630, 0.5923, 0.4992, 0.3936, 0.5844, 0.4975])
+        super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+
     def test_unidiffuser_default_joint_v0(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()