Fix kandinsky remove safety (#4065)

* Finish

* make style

src/diffusers/pipelines/kandinsky/pipeline_kandinsky.py

@@ -273,9 +273,6 @@ class KandinskyPipeline(DiffusionPipeline):
         for cpu_offloaded_model in [self.text_encoder, self.unet, self.movq]:
             _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
 
-        if self.safety_checker is not None:
-            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
-
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 

src/diffusers/pipelines/kandinsky/pipeline_kandinsky_img2img.py

@@ -308,9 +308,6 @@ class KandinskyImg2ImgPipeline(DiffusionPipeline):
         for cpu_offloaded_model in [self.text_encoder, self.unet, self.movq]:
             _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
 
-        if self.safety_checker is not None:
-            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
-
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 

src/diffusers/pipelines/kandinsky/pipeline_kandinsky_inpaint.py

@@ -433,9 +433,6 @@ class KandinskyInpaintPipeline(DiffusionPipeline):
         for cpu_offloaded_model in [self.text_encoder, self.unet, self.movq]:
             _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
 
-        if self.safety_checker is not None:
-            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
-
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 

tests/pipelines/kandinsky/test_kandinsky.py

@@ -230,6 +230,33 @@ class KandinskyPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2
         ), f" expected_slice {expected_slice}, but got {image_from_tuple_slice.flatten()}"
 
+    @require_torch_gpu
+    def test_offloads(self):
+        pipes = []
+        components = self.get_dummy_components()
+        sd_pipe = self.pipeline_class(**components).to(torch_device)
+        pipes.append(sd_pipe)
+
+        components = self.get_dummy_components()
+        sd_pipe = self.pipeline_class(**components)
+        sd_pipe.enable_model_cpu_offload()
+        pipes.append(sd_pipe)
+
+        components = self.get_dummy_components()
+        sd_pipe = self.pipeline_class(**components)
+        sd_pipe.enable_sequential_cpu_offload()
+        pipes.append(sd_pipe)
+
+        image_slices = []
+        for pipe in pipes:
+            inputs = self.get_dummy_inputs(torch_device)
+            image = pipe(**inputs).images
+
+            image_slices.append(image[0, -3:, -3:, -1].flatten())
+
+        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
+        assert np.abs(image_slices[0] - image_slices[2]).max() < 1e-3
+
 
 @slow
 @require_torch_gpu

tests/pipelines/kandinsky/test_kandinsky_img2img.py

@@ -242,6 +242,33 @@ class KandinskyImg2ImgPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2
         ), f" expected_slice {expected_slice}, but got {image_from_tuple_slice.flatten()}"
 
+    @require_torch_gpu
+    def test_offloads(self):
+        pipes = []
+        components = self.get_dummy_components()
+        sd_pipe = self.pipeline_class(**components).to(torch_device)
+        pipes.append(sd_pipe)
+
+        components = self.get_dummy_components()
+        sd_pipe = self.pipeline_class(**components)
+        sd_pipe.enable_model_cpu_offload()
+        pipes.append(sd_pipe)
+
+        components = self.get_dummy_components()
+        sd_pipe = self.pipeline_class(**components)
+        sd_pipe.enable_sequential_cpu_offload()
+        pipes.append(sd_pipe)
+
+        image_slices = []
+        for pipe in pipes:
+            inputs = self.get_dummy_inputs(torch_device)
+            image = pipe(**inputs).images
+
+            image_slices.append(image[0, -3:, -3:, -1].flatten())
+
+        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
+        assert np.abs(image_slices[0] - image_slices[2]).max() < 1e-3
+
 
 @slow
 @require_torch_gpu

tests/pipelines/kandinsky/test_kandinsky_inpaint.py

@@ -250,6 +250,33 @@ class KandinskyInpaintPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     def test_inference_batch_single_identical(self):
         super().test_inference_batch_single_identical(expected_max_diff=3e-3)
 
+    @require_torch_gpu
+    def test_offloads(self):
+        pipes = []
+        components = self.get_dummy_components()
+        sd_pipe = self.pipeline_class(**components).to(torch_device)
+        pipes.append(sd_pipe)
+
+        components = self.get_dummy_components()
+        sd_pipe = self.pipeline_class(**components)
+        sd_pipe.enable_model_cpu_offload()
+        pipes.append(sd_pipe)
+
+        components = self.get_dummy_components()
+        sd_pipe = self.pipeline_class(**components)
+        sd_pipe.enable_sequential_cpu_offload()
+        pipes.append(sd_pipe)
+
+        image_slices = []
+        for pipe in pipes:
+            inputs = self.get_dummy_inputs(torch_device)
+            image = pipe(**inputs).images
+
+            image_slices.append(image[0, -3:, -3:, -1].flatten())
+
+        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
+        assert np.abs(image_slices[0] - image_slices[2]).max() < 1e-3
+
 
 @slow
 @require_torch_gpu