update

* update

* update

* update

* update

.github/workflows/nightly_tests.yml

@@ -116,6 +116,7 @@ jobs:
       run:
         shell: bash
     strategy:
+      fail-fast: false
       max-parallel: 2
       matrix:
         module: [models, schedulers, lora, others, single_file, examples]

.github/workflows/push_tests.yml

@@ -112,6 +112,8 @@ jobs:
       run:
         shell: bash
     strategy:
+      fail-fast: false
+      max-parallel: 2
       matrix:
         module: [models, schedulers, lora, others, single_file]
     steps:

examples/textual_inversion/README.md

@@ -109,6 +109,9 @@ import torch
 model_id = "path-to-your-trained-model"
 pipe = StableDiffusionPipeline.from_pretrained(model_id,torch_dtype=torch.float16).to("cuda")
 
+repo_id_embeds = "path-to-your-learned-embeds"
+pipe.load_textual_inversion(repo_id_embeds)
+
 prompt = "A <cat-toy> backpack"
 
 image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]

src/diffusers/pipelines/pipeline_loading_utils.py

@@ -89,49 +89,44 @@ for library in LOADABLE_CLASSES:
     ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library])
 
 
-def is_safetensors_compatible(filenames, variant=None, passed_components=None) -> bool:
+def is_safetensors_compatible(filenames, passed_components=None) -> bool:
     """
     Checking for safetensors compatibility:
-    - By default, all models are saved with the default pytorch serialization, so we use the list of default pytorch
-      files to know which safetensors files are needed.
-    - The model is safetensors compatible only if there is a matching safetensors file for every default pytorch file.
+    - The model is safetensors compatible only if there is a safetensors file for each model component present in
+      filenames.
 
     Converting default pytorch serialized filenames to safetensors serialized filenames:
     - For models from the diffusers library, just replace the ".bin" extension with ".safetensors"
     - For models from the transformers library, the filename changes from "pytorch_model" to "model", and the ".bin"
       extension is replaced with ".safetensors"
     """
-    pt_filenames = []
-
-    sf_filenames = set()
-
     passed_components = passed_components or []
 
+    # extract all components of the pipeline and their associated files
+    components = {}
     for filename in filenames:
-        _, extension = os.path.splitext(filename)
-
-        if len(filename.split("/")) == 2 and filename.split("/")[0] in passed_components:
+        if not len(filename.split("/")) == 2:
             continue
 
-        if extension == ".bin":
-            pt_filenames.append(os.path.normpath(filename))
-        elif extension == ".safetensors":
-            sf_filenames.add(os.path.normpath(filename))
+        component, component_filename = filename.split("/")
+        if component in passed_components:
+            continue
 
-    for filename in pt_filenames:
-        #  filename = 'foo/bar/baz.bam' -> path = 'foo/bar', filename = 'baz', extension = '.bam'
-        path, filename = os.path.split(filename)
-        filename, extension = os.path.splitext(filename)
+        components.setdefault(component, [])
+        components[component].append(component_filename)
 
-        if filename.startswith("pytorch_model"):
-            filename = filename.replace("pytorch_model", "model")
-        else:
-            filename = filename
+    # iterate over all files of a component
+    # check if safetensor files exist for that component
+    # if variant is provided check if the variant of the safetensors exists
+    for component, component_filenames in components.items():
+        matches = []
+        for component_filename in component_filenames:
+            filename, extension = os.path.splitext(component_filename)
 
-        expected_sf_filename = os.path.normpath(os.path.join(path, filename))
-        expected_sf_filename = f"{expected_sf_filename}.safetensors"
-        if expected_sf_filename not in sf_filenames:
-            logger.warning(f"{expected_sf_filename} not found")
+            match_exists = extension == ".safetensors"
+            matches.append(match_exists)
+
+        if not any(matches):
             return False
 
     return True

src/diffusers/pipelines/pipeline_utils.py

@@ -1416,18 +1416,14 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             if (
                 use_safetensors
                 and not allow_pickle
-                and not is_safetensors_compatible(
-                    model_filenames, variant=variant, passed_components=passed_components
-                )
+                and not is_safetensors_compatible(model_filenames, passed_components=passed_components)
             ):
                 raise EnvironmentError(
                     f"Could not find the necessary `safetensors` weights in {model_filenames} (variant={variant})"
                 )
             if from_flax:
                 ignore_patterns = ["*.bin", "*.safetensors", "*.onnx", "*.pb"]
-            elif use_safetensors and is_safetensors_compatible(
-                model_filenames, variant=variant, passed_components=passed_components
-            ):
+            elif use_safetensors and is_safetensors_compatible(model_filenames, passed_components=passed_components):
                 ignore_patterns = ["*.bin", "*.msgpack"]
 
                 use_onnx = use_onnx if use_onnx is not None else pipeline_class._is_onnx

tests/lora/test_lora_layers_sd3.py

@@ -32,7 +32,7 @@ from utils import PeftLoraLoaderMixinTests  # noqa: E402
 @require_peft_backend
 class SD3LoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = StableDiffusion3Pipeline
-    scheduler_cls = FlowMatchEulerDiscreteScheduler()
+    scheduler_cls = FlowMatchEulerDiscreteScheduler
     scheduler_kwargs = {}
     uses_flow_matching = True
     transformer_kwargs = {
@@ -80,8 +80,7 @@ class SD3LoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
         Related PR: https://github.com/huggingface/diffusers/pull/8584
         """
         components = self.get_dummy_components()
-
-        pipe = self.pipeline_class(**components)
+        pipe = self.pipeline_class(**components[0])
         pipe = pipe.to(torch_device)
         pipe.set_progress_bar_config(disable=None)
 

tests/lora/test_lora_layers_sdxl.py

@@ -124,71 +124,6 @@ class LoraSDXLIntegrationTests(unittest.TestCase):
         gc.collect()
         torch.cuda.empty_cache()
 
-    def test_sdxl_0_9_lora_one(self):
-        generator = torch.Generator().manual_seed(0)
-
-        pipe = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9")
-        lora_model_id = "hf-internal-testing/sdxl-0.9-daiton-lora"
-        lora_filename = "daiton-xl-lora-test.safetensors"
-        pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)
-        pipe.enable_model_cpu_offload()
-
-        images = pipe(
-            "masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
-        ).images
-
-        images = images[0, -3:, -3:, -1].flatten()
-        expected = np.array([0.3838, 0.3482, 0.3588, 0.3162, 0.319, 0.3369, 0.338, 0.3366, 0.3213])
-
-        max_diff = numpy_cosine_similarity_distance(expected, images)
-        assert max_diff < 1e-3
-        pipe.unload_lora_weights()
-        release_memory(pipe)
-
-    def test_sdxl_0_9_lora_two(self):
-        generator = torch.Generator().manual_seed(0)
-
-        pipe = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9")
-        lora_model_id = "hf-internal-testing/sdxl-0.9-costumes-lora"
-        lora_filename = "saijo.safetensors"
-        pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)
-        pipe.enable_model_cpu_offload()
-
-        images = pipe(
-            "masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
-        ).images
-
-        images = images[0, -3:, -3:, -1].flatten()
-        expected = np.array([0.3137, 0.3269, 0.3355, 0.255, 0.2577, 0.2563, 0.2679, 0.2758, 0.2626])
-
-        max_diff = numpy_cosine_similarity_distance(expected, images)
-        assert max_diff < 1e-3
-
-        pipe.unload_lora_weights()
-        release_memory(pipe)
-
-    def test_sdxl_0_9_lora_three(self):
-        generator = torch.Generator().manual_seed(0)
-
-        pipe = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9")
-        lora_model_id = "hf-internal-testing/sdxl-0.9-kamepan-lora"
-        lora_filename = "kame_sdxl_v2-000020-16rank.safetensors"
-        pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)
-        pipe.enable_model_cpu_offload()
-
-        images = pipe(
-            "masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
-        ).images
-
-        images = images[0, -3:, -3:, -1].flatten()
-        expected = np.array([0.4015, 0.3761, 0.3616, 0.3745, 0.3462, 0.3337, 0.3564, 0.3649, 0.3468])
-
-        max_diff = numpy_cosine_similarity_distance(expected, images)
-        assert max_diff < 5e-3
-
-        pipe.unload_lora_weights()
-        release_memory(pipe)
-
     def test_sdxl_1_0_lora(self):
         generator = torch.Generator("cpu").manual_seed(0)
 

tests/models/transformers/test_models_transformer_aura_flow.py

@@ -26,7 +26,7 @@ from ..test_modeling_common import ModelTesterMixin
 enable_full_determinism()
 
 
-class SD3TransformerTests(ModelTesterMixin, unittest.TestCase):
+class AuraFlowTransformerTests(ModelTesterMixin, unittest.TestCase):
     model_class = AuraFlowTransformer2DModel
     main_input_name = "hidden_states"
     # We override the items here because the transformer under consideration is small.
@@ -73,3 +73,7 @@ class SD3TransformerTests(ModelTesterMixin, unittest.TestCase):
         }
         inputs_dict = self.dummy_input
         return init_dict, inputs_dict
+
+    @unittest.skip("AuraFlowTransformer2DModel uses its own dedicated attention processor. This test does not apply")
+    def test_set_attn_processor_for_determinism(self):
+        pass

tests/models/transformers/test_models_transformer_sd3.py

@@ -76,3 +76,7 @@ class SD3TransformerTests(ModelTesterMixin, unittest.TestCase):
         }
         inputs_dict = self.dummy_input
         return init_dict, inputs_dict
+
+    @unittest.skip("SD3Transformer2DModel uses a dedicated attention processor. This test doesn't apply")
+    def test_set_attn_processor_for_determinism(self):
+        pass

tests/pipelines/aura_flow/test_pipeline_aura_flow.py

@@ -163,3 +163,7 @@ class AuraFlowPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
         assert np.allclose(
             original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
         ), "Original outputs should match when fused QKV projections are disabled."
+
+    @unittest.skip("xformers attention processor does not exist for AuraFlow")
+    def test_xformers_attention_forwardGenerator_pass(self):
+        pass

tests/pipelines/lumina/test_lumina_nextdit.py

@@ -119,6 +119,10 @@ class LuminaText2ImgPipelinePipelineFastTests(unittest.TestCase, PipelineTesterM
         max_diff = np.abs(output_with_prompt - output_with_embeds).max()
         assert max_diff < 1e-4
 
+    @unittest.skip("xformers attention processor does not exist for Lumina")
+    def test_xformers_attention_forwardGenerator_pass(self):
+        pass
+
 
 @slow
 @require_torch_gpu

tests/pipelines/test_pipeline_utils.py

@@ -68,25 +68,21 @@ class IsSafetensorsCompatibleTests(unittest.TestCase):
             "unet/diffusion_pytorch_model.fp16.bin",
             "unet/diffusion_pytorch_model.fp16.safetensors",
         ]
-        variant = "fp16"
-        self.assertTrue(is_safetensors_compatible(filenames, variant=variant))
+        self.assertTrue(is_safetensors_compatible(filenames))
 
     def test_diffusers_model_is_compatible_variant(self):
         filenames = [
             "unet/diffusion_pytorch_model.fp16.bin",
             "unet/diffusion_pytorch_model.fp16.safetensors",
         ]
-        variant = "fp16"
-        self.assertTrue(is_safetensors_compatible(filenames, variant=variant))
+        self.assertTrue(is_safetensors_compatible(filenames))
 
-    def test_diffusers_model_is_compatible_variant_partial(self):
-        # pass variant but use the non-variant filenames
+    def test_diffusers_model_is_compatible_variant_mixed(self):
         filenames = [
             "unet/diffusion_pytorch_model.bin",
-            "unet/diffusion_pytorch_model.safetensors",
+            "unet/diffusion_pytorch_model.fp16.safetensors",
         ]
-        variant = "fp16"
-        self.assertTrue(is_safetensors_compatible(filenames, variant=variant))
+        self.assertTrue(is_safetensors_compatible(filenames))
 
     def test_diffusers_model_is_not_compatible_variant(self):
         filenames = [
@@ -99,25 +95,14 @@ class IsSafetensorsCompatibleTests(unittest.TestCase):
             "unet/diffusion_pytorch_model.fp16.bin",
             # Removed: 'unet/diffusion_pytorch_model.fp16.safetensors',
         ]
-        variant = "fp16"
-        self.assertFalse(is_safetensors_compatible(filenames, variant=variant))
+        self.assertFalse(is_safetensors_compatible(filenames))
 
     def test_transformer_model_is_compatible_variant(self):
         filenames = [
             "text_encoder/pytorch_model.fp16.bin",
             "text_encoder/model.fp16.safetensors",
         ]
-        variant = "fp16"
-        self.assertTrue(is_safetensors_compatible(filenames, variant=variant))
-
-    def test_transformer_model_is_compatible_variant_partial(self):
-        # pass variant but use the non-variant filenames
-        filenames = [
-            "text_encoder/pytorch_model.bin",
-            "text_encoder/model.safetensors",
-        ]
-        variant = "fp16"
-        self.assertTrue(is_safetensors_compatible(filenames, variant=variant))
+        self.assertTrue(is_safetensors_compatible(filenames))
 
     def test_transformer_model_is_not_compatible_variant(self):
         filenames = [
@@ -126,9 +111,45 @@ class IsSafetensorsCompatibleTests(unittest.TestCase):
             "vae/diffusion_pytorch_model.fp16.bin",
             "vae/diffusion_pytorch_model.fp16.safetensors",
             "text_encoder/pytorch_model.fp16.bin",
-            # 'text_encoder/model.fp16.safetensors',
             "unet/diffusion_pytorch_model.fp16.bin",
             "unet/diffusion_pytorch_model.fp16.safetensors",
         ]
-        variant = "fp16"
-        self.assertFalse(is_safetensors_compatible(filenames, variant=variant))
+        self.assertFalse(is_safetensors_compatible(filenames))
+
+    def test_transformers_is_compatible_sharded(self):
+        filenames = [
+            "text_encoder/pytorch_model.bin",
+            "text_encoder/model-00001-of-00002.safetensors",
+            "text_encoder/model-00002-of-00002.safetensors",
+        ]
+        self.assertTrue(is_safetensors_compatible(filenames))
+
+    def test_transformers_is_compatible_variant_sharded(self):
+        filenames = [
+            "text_encoder/pytorch_model.bin",
+            "text_encoder/model.fp16-00001-of-00002.safetensors",
+            "text_encoder/model.fp16-00001-of-00002.safetensors",
+        ]
+        self.assertTrue(is_safetensors_compatible(filenames))
+
+    def test_diffusers_is_compatible_sharded(self):
+        filenames = [
+            "unet/diffusion_pytorch_model.bin",
+            "unet/diffusion_pytorch_model-00001-of-00002.safetensors",
+            "unet/diffusion_pytorch_model-00002-of-00002.safetensors",
+        ]
+        self.assertTrue(is_safetensors_compatible(filenames))
+
+    def test_diffusers_is_compatible_variant_sharded(self):
+        filenames = [
+            "unet/diffusion_pytorch_model.bin",
+            "unet/diffusion_pytorch_model.fp16-00001-of-00002.safetensors",
+            "unet/diffusion_pytorch_model.fp16-00001-of-00002.safetensors",
+        ]
+        self.assertTrue(is_safetensors_compatible(filenames))
+
+    def test_diffusers_is_compatible_only_variants(self):
+        filenames = [
+            "unet/diffusion_pytorch_model.fp16.safetensors",
+        ]
+        self.assertTrue(is_safetensors_compatible(filenames))

tests/pipelines/test_pipelines.py

@@ -551,37 +551,94 @@ class DownloadTests(unittest.TestCase):
                 assert sum(f.endswith(this_format) and not f.endswith(f"{variant}{this_format}") for f in files) == 3
                 assert not any(f.endswith(other_format) for f in files)
 
-    def test_download_broken_variant(self):
-        for use_safetensors in [False, True]:
-            # text encoder is missing no variant and "no_ema" variant weights, so the following can't work
-            for variant in [None, "no_ema"]:
-                with self.assertRaises(OSError) as error_context:
-                    with tempfile.TemporaryDirectory() as tmpdirname:
-                        tmpdirname = StableDiffusionPipeline.from_pretrained(
-                            "hf-internal-testing/stable-diffusion-broken-variants",
-                            cache_dir=tmpdirname,
-                            variant=variant,
-                            use_safetensors=use_safetensors,
-                        )
+    def test_download_safetensors_only_variant_exists_for_model(self):
+        variant = None
+        use_safetensors = True
 
-                assert "Error no file name" in str(error_context.exception)
-
-            # text encoder has fp16 variants so we can load it
-            with tempfile.TemporaryDirectory() as tmpdirname:
-                tmpdirname = StableDiffusionPipeline.download(
+        # text encoder is missing no variant weights, so the following can't work
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            with self.assertRaises(OSError) as error_context:
+                tmpdirname = StableDiffusionPipeline.from_pretrained(
                     "hf-internal-testing/stable-diffusion-broken-variants",
-                    use_safetensors=use_safetensors,
                     cache_dir=tmpdirname,
-                    variant="fp16",
+                    variant=variant,
+                    use_safetensors=use_safetensors,
+                )
+            assert "Error no file name" in str(error_context.exception)
+
+        # text encoder has fp16 variants so we can load it
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            tmpdirname = StableDiffusionPipeline.download(
+                "hf-internal-testing/stable-diffusion-broken-variants",
+                use_safetensors=use_safetensors,
+                cache_dir=tmpdirname,
+                variant="fp16",
+            )
+            all_root_files = [t[-1] for t in os.walk(tmpdirname)]
+            files = [item for sublist in all_root_files for item in sublist]
+            # None of the downloaded files should be a non-variant file even if we have some here:
+            # https://huggingface.co/hf-internal-testing/stable-diffusion-broken-variants/tree/main/unet
+            assert len(files) == 15, f"We should only download 15 files, not {len(files)}"
+
+    def test_download_bin_only_variant_exists_for_model(self):
+        variant = None
+        use_safetensors = False
+
+        # text encoder is missing Non-variant weights, so the following can't work
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            with self.assertRaises(OSError) as error_context:
+                tmpdirname = StableDiffusionPipeline.from_pretrained(
+                    "hf-internal-testing/stable-diffusion-broken-variants",
+                    cache_dir=tmpdirname,
+                    variant=variant,
+                    use_safetensors=use_safetensors,
+                )
+            assert "Error no file name" in str(error_context.exception)
+
+        # text encoder has fp16 variants so we can load it
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            tmpdirname = StableDiffusionPipeline.download(
+                "hf-internal-testing/stable-diffusion-broken-variants",
+                use_safetensors=use_safetensors,
+                cache_dir=tmpdirname,
+                variant="fp16",
+            )
+            all_root_files = [t[-1] for t in os.walk(tmpdirname)]
+            files = [item for sublist in all_root_files for item in sublist]
+            # None of the downloaded files should be a non-variant file even if we have some here:
+            # https://huggingface.co/hf-internal-testing/stable-diffusion-broken-variants/tree/main/unet
+            assert len(files) == 15, f"We should only download 15 files, not {len(files)}"
+
+    def test_download_safetensors_variant_does_not_exist_for_model(self):
+        variant = "no_ema"
+        use_safetensors = True
+
+        # text encoder is missing no_ema variant weights, so the following can't work
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            with self.assertRaises(OSError) as error_context:
+                tmpdirname = StableDiffusionPipeline.from_pretrained(
+                    "hf-internal-testing/stable-diffusion-broken-variants",
+                    cache_dir=tmpdirname,
+                    variant=variant,
+                    use_safetensors=use_safetensors,
                 )
 
-                all_root_files = [t[-1] for t in os.walk(tmpdirname)]
-                files = [item for sublist in all_root_files for item in sublist]
+            assert "Error no file name" in str(error_context.exception)
 
-                # None of the downloaded files should be a non-variant file even if we have some here:
-                # https://huggingface.co/hf-internal-testing/stable-diffusion-broken-variants/tree/main/unet
-                assert len(files) == 15, f"We should only download 15 files, not {len(files)}"
-                # only unet has "no_ema" variant
+    def test_download_bin_variant_does_not_exist_for_model(self):
+        variant = "no_ema"
+        use_safetensors = False
+
+        # text encoder is missing no_ema variant weights, so the following can't work
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            with self.assertRaises(OSError) as error_context:
+                tmpdirname = StableDiffusionPipeline.from_pretrained(
+                    "hf-internal-testing/stable-diffusion-broken-variants",
+                    cache_dir=tmpdirname,
+                    variant=variant,
+                    use_safetensors=use_safetensors,
+                )
+            assert "Error no file name" in str(error_context.exception)
 
     def test_local_save_load_index(self):
         prompt = "hello"

tests/pipelines/text_to_video_synthesis/test_text_to_video.py

@@ -20,12 +20,7 @@ import numpy as np
 import torch
 from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
 
-from diffusers import (
-    AutoencoderKL,
-    DDIMScheduler,
-    TextToVideoSDPipeline,
-    UNet3DConditionModel,
-)
+from diffusers import AutoencoderKL, DDIMScheduler, TextToVideoSDPipeline, UNet3DConditionModel
 from diffusers.utils import is_xformers_available
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
@@ -64,7 +59,7 @@ class TextToVideoSDPipelineFastTests(PipelineTesterMixin, SDFunctionTesterMixin,
     def get_dummy_components(self):
         torch.manual_seed(0)
         unet = UNet3DConditionModel(
-            block_out_channels=(4, 8),
+            block_out_channels=(8, 8),
             layers_per_block=1,
             sample_size=32,
             in_channels=4,
@@ -134,10 +129,7 @@ class TextToVideoSDPipelineFastTests(PipelineTesterMixin, SDFunctionTesterMixin,
         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)
+        return super().test_dict_tuple_outputs_equivalent()
 
     def test_text_to_video_default_case(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
@@ -151,9 +143,8 @@ class TextToVideoSDPipelineFastTests(PipelineTesterMixin, SDFunctionTesterMixin,
         frames = sd_pipe(**inputs).frames
 
         image_slice = frames[0][0][-3:, -3:, -1]
-
         assert frames[0][0].shape == (32, 32, 3)
-        expected_slice = np.array([0.7537, 0.1752, 0.6157, 0.5508, 0.4240, 0.4110, 0.4838, 0.5648, 0.5094])
+        expected_slice = np.array([0.8093, 0.2751, 0.6976, 0.5927, 0.4616, 0.4336, 0.5094, 0.5683, 0.4796])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2