Merge branch 'main' into tests/add-test-slices-wan

Author: a-r-r-o-w

examples/dreambooth/train_dreambooth_lora_flux_kontext.py

@@ -1614,7 +1614,7 @@ def load_model_hook(models, input_dir):
     )
     if args.cond_image_column is not None:
         logger.info("I2I fine-tuning enabled.")
-    batch_sampler = BucketBatchSampler(train_dataset, batch_size=args.train_batch_size, drop_last=False)
+    batch_sampler = BucketBatchSampler(train_dataset, batch_size=args.train_batch_size, drop_last=True)
     train_dataloader = torch.utils.data.DataLoader(
         train_dataset,
         batch_sampler=batch_sampler,

tests/pipelines/cosmos/test_cosmos.py

@@ -153,11 +153,15 @@ def test_inference(self):
         inputs = self.get_dummy_inputs(device)
         video = pipe(**inputs).frames
         generated_video = video[0]
-
         self.assertEqual(generated_video.shape, (9, 3, 32, 32))
-        expected_video = torch.randn(9, 3, 32, 32)
-        max_diff = np.abs(generated_video - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.0, 0.9686, 0.8549, 0.8078, 0.0, 0.8431, 1.0, 0.4863, 0.7098, 0.1098, 0.8157, 0.4235, 0.6353, 0.2549, 0.5137, 0.5333])
+        # fmt: on
+
+        generated_slice = generated_video.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))
 
     def test_callback_inputs(self):
         sig = inspect.signature(self.pipeline_class.__call__)

tests/pipelines/cosmos/test_cosmos2_text2image.py

@@ -140,11 +140,15 @@ def test_inference(self):
         inputs = self.get_dummy_inputs(device)
         image = pipe(**inputs).images
         generated_image = image[0]
-
         self.assertEqual(generated_image.shape, (3, 32, 32))
-        expected_video = torch.randn(3, 32, 32)
-        max_diff = np.abs(generated_image - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.451, 0.451, 0.4471, 0.451, 0.451, 0.451, 0.451, 0.451, 0.4784, 0.4784, 0.4784, 0.4784, 0.4784, 0.4902, 0.4588, 0.5333])
+        # fmt: on
+
+        generated_slice = generated_image.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))
 
     def test_callback_inputs(self):
         sig = inspect.signature(self.pipeline_class.__call__)

tests/pipelines/cosmos/test_cosmos2_video2world.py

@@ -147,11 +147,15 @@ def test_inference(self):
         inputs = self.get_dummy_inputs(device)
         video = pipe(**inputs).frames
         generated_video = video[0]
-
         self.assertEqual(generated_video.shape, (9, 3, 32, 32))
-        expected_video = torch.randn(9, 3, 32, 32)
-        max_diff = np.abs(generated_video - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.451, 0.451, 0.4471, 0.451, 0.451, 0.451, 0.451, 0.451, 0.5098, 0.5137, 0.5176, 0.5098, 0.5255, 0.5412, 0.5098, 0.5059])
+        # fmt: on
+
+        generated_slice = generated_video.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))
 
     def test_components_function(self):
         init_components = self.get_dummy_components()

tests/pipelines/cosmos/test_cosmos_video2world.py

@@ -159,11 +159,15 @@ def test_inference(self):
         inputs = self.get_dummy_inputs(device)
         video = pipe(**inputs).frames
         generated_video = video[0]
-
         self.assertEqual(generated_video.shape, (9, 3, 32, 32))
-        expected_video = torch.randn(9, 3, 32, 32)
-        max_diff = np.abs(generated_video - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.0, 0.8275, 0.7529, 0.7294, 0.0, 0.6, 1.0, 0.3804, 0.6667, 0.0863, 0.8784, 0.5922, 0.6627, 0.2784, 0.5725, 0.7765])
+        # fmt: on
+
+        generated_slice = generated_video.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))
 
     def test_components_function(self):
         init_components = self.get_dummy_components()

tests/pipelines/hidream_image/test_pipeline_hidream.py

@@ -146,11 +146,15 @@ def test_inference(self):
         inputs = self.get_dummy_inputs(device)
         image = pipe(**inputs)[0]
         generated_image = image[0]
-
         self.assertEqual(generated_image.shape, (128, 128, 3))
-        expected_image = torch.randn(128, 128, 3).numpy()
-        max_diff = np.abs(generated_image - expected_image).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = np.array([0.4507, 0.5256, 0.4205, 0.5791, 0.4848, 0.4831, 0.4443, 0.5107, 0.6586, 0.3163, 0.7318, 0.5933, 0.6252, 0.5512, 0.5357, 0.5983])
+        # fmt: on
+
+        generated_slice = generated_image.flatten()
+        generated_slice = np.concatenate([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(np.allclose(generated_slice, expected_slice, atol=1e-3))
 
     def test_inference_batch_single_identical(self):
         super().test_inference_batch_single_identical(expected_max_diff=3e-4)

tests/pipelines/hunyuan_video/test_hunyuan_image2video.py

@@ -229,12 +229,19 @@ def test_inference(self):
         inputs = self.get_dummy_inputs(device)
         video = pipe(**inputs).frames
         generated_video = video[0]
-
         # NOTE: The expected video has 4 lesser frames because they are dropped in the pipeline
         self.assertEqual(generated_video.shape, (5, 3, 16, 16))
-        expected_video = torch.randn(5, 3, 16, 16)
-        max_diff = np.abs(generated_video - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.444, 0.479, 0.4485, 0.5752, 0.3539, 0.1548, 0.2706, 0.3593, 0.5323, 0.6635, 0.6795, 0.5255, 0.5091, 0.345, 0.4276, 0.4128])
+        # fmt: on
+
+        generated_slice = generated_video.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(
+            torch.allclose(generated_slice, expected_slice, atol=1e-3),
+            "The generated video does not match the expected slice.",
+        )
 
     def test_callback_inputs(self):
         sig = inspect.signature(self.pipeline_class.__call__)

tests/pipelines/hunyuan_video/test_hunyuan_skyreels_image2video.py

@@ -192,11 +192,18 @@ def test_inference(self):
         inputs = self.get_dummy_inputs(device)
         video = pipe(**inputs).frames
         generated_video = video[0]
-
         self.assertEqual(generated_video.shape, (9, 3, 16, 16))
-        expected_video = torch.randn(9, 3, 16, 16)
-        max_diff = np.abs(generated_video - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.5832, 0.5498, 0.4839, 0.4744, 0.4515, 0.4832, 0.496, 0.563, 0.5918, 0.5979, 0.5101, 0.6168, 0.6613, 0.536, 0.55, 0.5775])
+        # fmt: on
+
+        generated_slice = generated_video.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(
+            torch.allclose(generated_slice, expected_slice, atol=1e-3),
+            "The generated video does not match the expected slice.",
+        )
 
     def test_callback_inputs(self):
         sig = inspect.signature(self.pipeline_class.__call__)

tests/pipelines/hunyuan_video/test_hunyuan_video.py

@@ -26,10 +26,7 @@
     HunyuanVideoPipeline,
     HunyuanVideoTransformer3DModel,
 )
-from diffusers.utils.testing_utils import (
-    enable_full_determinism,
-    torch_device,
-)
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
 
 from ..test_pipelines_common import (
     FasterCacheTesterMixin,
@@ -206,11 +203,18 @@ def test_inference(self):
         inputs = self.get_dummy_inputs(device)
         video = pipe(**inputs).frames
         generated_video = video[0]
-
         self.assertEqual(generated_video.shape, (9, 3, 16, 16))
-        expected_video = torch.randn(9, 3, 16, 16)
-        max_diff = np.abs(generated_video - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.3946, 0.4649, 0.3196, 0.4569, 0.3312, 0.3687, 0.3216, 0.3972, 0.4469, 0.3888, 0.3929, 0.3802, 0.3479, 0.3888, 0.3825, 0.3542])
+        # fmt: on
+
+        generated_slice = generated_video.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(
+            torch.allclose(generated_slice, expected_slice, atol=1e-3),
+            "The generated video does not match the expected slice.",
+        )
 
     def test_callback_inputs(self):
         sig = inspect.signature(self.pipeline_class.__call__)

tests/pipelines/hunyuan_video/test_hunyuan_video_framepack.py

@@ -227,11 +227,18 @@ def test_inference(self):
         inputs = self.get_dummy_inputs(device)
         video = pipe(**inputs).frames
         generated_video = video[0]
-
         self.assertEqual(generated_video.shape, (13, 3, 32, 32))
-        expected_video = torch.randn(13, 3, 32, 32)
-        max_diff = np.abs(generated_video - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.363, 0.3384, 0.3426, 0.3512, 0.3372, 0.3276, 0.417, 0.4061, 0.5221, 0.467, 0.4813, 0.4556, 0.4107, 0.3945, 0.4049, 0.4551])
+        # fmt: on
+
+        generated_slice = generated_video.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(
+            torch.allclose(generated_slice, expected_slice, atol=1e-3),
+            "The generated video does not match the expected slice.",
+        )
 
     def test_callback_inputs(self):
         sig = inspect.signature(self.pipeline_class.__call__)