Merge branch 'main' into to-single-file/wan

Author: a-r-r-o-w

docs/source/en/_toctree.yml

@@ -239,6 +239,12 @@ The `step()` function is [called](https://github.com/huggingface/diffusers/blob/
 
 In general, the `sigmas` should [stay on the CPU](https://github.com/huggingface/diffusers/blob/35a969d297cba69110d175ee79c59312b9f49e1e/src/diffusers/schedulers/scheduling_euler_discrete.py#L240) to avoid the communication sync and latency.
 
+<Tip>
+
+Refer to the [torch.compile and Diffusers: A Hands-On Guide to Peak Performance](https://pytorch.org/blog/torch-compile-and-diffusers-a-hands-on-guide-to-peak-performance/) blog post for maximizing performance with `torch.compile` for diffusion models.
+
+</Tip>
+
 ### Benchmarks
 
 Refer to the [diffusers/benchmarks](https://huggingface.co/datasets/diffusers/benchmarks) dataset to see inference latency and memory usage data for compiled pipelines.
@@ -298,4 +304,6 @@ pipeline.fuse_qkv_projections()
 
 - Read the [Presenting Flux Fast: Making Flux go brrr on H100s](https://pytorch.org/blog/presenting-flux-fast-making-flux-go-brrr-on-h100s/) blog post to learn more about how you can combine all of these optimizations with [TorchInductor](https://docs.pytorch.org/docs/stable/torch.compiler.html) and [AOTInductor](https://docs.pytorch.org/docs/stable/torch.compiler_aot_inductor.html) for a ~2.5x speedup using recipes from [flux-fast](https://github.com/huggingface/flux-fast).
 
-    These recipes support AMD hardware and [Flux.1 Kontext Dev](https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev).
+    These recipes support AMD hardware and [Flux.1 Kontext Dev](https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev).
+- Read the [torch.compile and Diffusers: A Hands-On Guide to Peak Performance](https://pytorch.org/blog/torch-compile-and-diffusers-a-hands-on-guide-to-peak-performance/) blog post
+to maximize performance when using `torch.compile`.

docs/source/en/optimization/fp16.md

@@ -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,

docs/source/en/tutorials/tutorial_overview.md

@@ -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__)

docs/source/en/using-diffusers/overview_techniques.md

@@ -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__)

examples/dreambooth/train_dreambooth_lora_flux_kontext.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.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/cosmos/test_cosmos2_text2image.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)