quick fix to make sure sdxl conversion works

Signed-off-by: Raphael <oOraph@users.noreply.github.com>
Co-authored-by: Raphael <oOraph@users.noreply.github.com>

docs/source/en/api/pipelines/stable_diffusion/stable_diffusion_xl.mdx

@@ -21,7 +21,7 @@ The abstract of the paper is the following:
 ## Tips
 
 - Stable Diffusion XL works especially well with images between 768 and 1024.
-- Stable Diffusion XL output image can be improved by making use of a refiner as shown below
+- Stable Diffusion XL output image can be improved by making use of a refiner as shown below.
 
 ### Available checkpoints:
 
@@ -40,7 +40,7 @@ pip install safetensors
 pip install invisible-watermark>=2.0
 ```
 
-### *Text-to-Image*
+### Text-to-Image
 
 You can use SDXL as follows for *text-to-image*:
 
@@ -71,6 +71,7 @@ pipe = StableDiffusionXLPipeline.from_pretrained(
 )
 pipe.to("cuda")
 
+use_refiner = True
 refiner = StableDiffusionXLImg2ImgPipeline.from_pretrained(
     "stabilityai/stable-diffusion-xl-refiner-0.9", torch_dtype=torch.float16, use_safetensors=True, variant="fp16"
 )
@@ -82,7 +83,29 @@ image = pipe(prompt=prompt, output_type="latent" if use_refiner else "pil").imag
 image = refiner(prompt=prompt, image=image[None, :]).images[0]
 ```
 
-### Loading single file checkpoitns / original file format
+### Image-to-image 
+
+```py 
+import torch
+from diffusers import StableDiffusionXLImg2ImgPipeline
+from diffusers.utils import load_image
+
+pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-refiner-0.9", torch_dtype=torch.float16
+)
+pipe = pipe.to("cuda")
+url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png"
+
+init_image = load_image(url).convert("RGB")
+prompt = "a photo of an astronaut riding a horse on mars"
+image = pipe(prompt, image=init_image).images[0]
+```
+
+| Original Image | Refined Image |
+|---|---|
+| ![](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/sd_xl/init_image.png) | ![](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/sd_xl/refined_image.png) |
+
+### Loading single file checkpoints / original file format
 
 By making use of [`~diffusers.loaders.FromSingleFileMixin.from_single_file`] you can also load the 
 original file format into `diffusers`:
@@ -127,7 +150,7 @@ You can speed up inference by making use of `torch.compile`. This should give yo
 + refiner.unet = torch.compile(refiner.unet, mode="reduce-overhead", fullgraph=True)
 ```
 
-### Running with `torch` < 2.0
+### Running with `torch` \< 2.0
 
 **Note** that if you want to run Stable Diffusion XL with `torch` < 2.0, please make sure to enable xformers 
 attention:

examples/controlnet/train_controlnet.py

@@ -56,7 +56,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__)
 

examples/controlnet/train_controlnet_flax.py

@@ -59,7 +59,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = logging.getLogger(__name__)
 

examples/custom_diffusion/train_custom_diffusion.py

@@ -57,7 +57,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__)
 

examples/dreambooth/train_dreambooth.py

@@ -59,7 +59,7 @@ if is_wandb_available():
     import wandb
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__)
 

examples/dreambooth/train_dreambooth_flax.py

@@ -36,7 +36,7 @@ from diffusers.utils import check_min_version
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 # Cache compiled models across invocations of this script.
 cc.initialize_cache(os.path.expanduser("~/.cache/jax/compilation_cache"))

examples/dreambooth/train_dreambooth_lora.py

@@ -65,7 +65,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__)
 

examples/instruct_pix2pix/train_instruct_pix2pix.py

@@ -52,7 +52,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/text_to_image/train_text_to_image.py

@@ -54,7 +54,7 @@ if is_wandb_available():
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/text_to_image/train_text_to_image_flax.py

@@ -33,7 +33,7 @@ from diffusers.utils import check_min_version
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = logging.getLogger(__name__)
 

examples/text_to_image/train_text_to_image_lora.py

@@ -48,7 +48,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__, log_level="INFO")
 

examples/textual_inversion/textual_inversion.py

@@ -78,7 +78,7 @@ else:
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__)
 

examples/textual_inversion/textual_inversion_flax.py

@@ -56,7 +56,7 @@ else:
 # ------------------------------------------------------------------------------
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = logging.getLogger(__name__)
 

examples/unconditional_image_generation/train_unconditional.py

@@ -29,7 +29,7 @@ from diffusers.utils.import_utils import is_xformers_available
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.18.0.dev0")
+check_min_version("0.18.0")
 
 logger = get_logger(__name__, log_level="INFO")
 

setup.py

@@ -232,7 +232,7 @@ install_requires = [
 
 setup(
     name="diffusers",
-    version="0.18.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="0.18.2",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     description="Diffusers",
     long_description=open("README.md", "r", encoding="utf-8").read(),
     long_description_content_type="text/markdown",

src/diffusers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.18.0.dev0"
+__version__ = "0.18.2"
 
 from .configuration_utils import ConfigMixin
 from .utils import (

src/diffusers/configuration_utils.py

@@ -607,7 +607,7 @@ def register_to_config(init):
 
         # Take note of the parameters that were not present in the loaded config
         if len(set(new_kwargs.keys()) - set(init_kwargs)) > 0:
-            new_kwargs["_use_default_values"] = set(new_kwargs.keys()) - set(init_kwargs)
+            new_kwargs["_use_default_values"] = list(set(new_kwargs.keys()) - set(init_kwargs))
 
         new_kwargs = {**config_init_kwargs, **new_kwargs}
         getattr(self, "register_to_config")(**new_kwargs)
@@ -655,7 +655,7 @@ def flax_register_to_config(cls):
 
         # Take note of the parameters that were not present in the loaded config
         if len(set(new_kwargs.keys()) - set(init_kwargs)) > 0:
-            new_kwargs["_use_default_values"] = set(new_kwargs.keys()) - set(init_kwargs)
+            new_kwargs["_use_default_values"] = list(set(new_kwargs.keys()) - set(init_kwargs))
 
         getattr(self, "register_to_config")(**new_kwargs)
         original_init(self, *args, **kwargs)

src/diffusers/loaders.py

@@ -1394,7 +1394,7 @@ class FromSingleFileMixin:
         use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         extract_ema = kwargs.pop("extract_ema", False)
-        image_size = kwargs.pop("image_size", 512)
+        image_size = kwargs.pop("image_size", None)
         scheduler_type = kwargs.pop("scheduler_type", "pndm")
         num_in_channels = kwargs.pop("num_in_channels", None)
         upcast_attention = kwargs.pop("upcast_attention", None)

src/diffusers/pipelines/pipeline_utils.py

@@ -1213,6 +1213,15 @@ class DiffusionPipeline(ConfigMixin):
             filenames = {sibling.rfilename for sibling in info.siblings}
             model_filenames, variant_filenames = variant_compatible_siblings(filenames, variant=variant)
 
+            if len(variant_filenames) == 0 and variant is not None:
+                deprecation_message = (
+                    f"You are trying to load the model files of the `variant={variant}`, but no such modeling files are available."
+                    f"The default model files: {model_filenames} will be loaded instead. Make sure to not load from `variant={variant}`"
+                    "if such variant modeling files are not available. Doing so will lead to an error in v0.22.0 as defaulting to non-variant"
+                    "modeling files is deprecated."
+                )
+                deprecate("no variant default", "0.22.0", deprecation_message, standard_warn=False)
+
             # remove ignored filenames
             model_filenames = set(model_filenames) - set(ignore_filenames)
             variant_filenames = set(variant_filenames) - set(ignore_filenames)
@@ -1302,7 +1311,7 @@ class DiffusionPipeline(ConfigMixin):
             snapshot_folder = Path(config_file).parent
             pipeline_is_cached = all((snapshot_folder / f).is_file() for f in expected_files)
 
-            if pipeline_is_cached:
+            if pipeline_is_cached and not force_download:
                 # if the pipeline is cached, we can directly return it
                 # else call snapshot_download
                 return snapshot_folder

src/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py

@@ -24,6 +24,7 @@ from transformers import (
     AutoFeatureExtractor,
     BertTokenizerFast,
     CLIPImageProcessor,
+    CLIPTextConfig,
     CLIPTextModel,
     CLIPTextModelWithProjection,
     CLIPTokenizer,
@@ -48,7 +49,7 @@ from ...schedulers import (
     PNDMScheduler,
     UnCLIPScheduler,
 )
-from ...utils import is_omegaconf_available, is_safetensors_available, logging
+from ...utils import is_accelerate_available, is_omegaconf_available, is_safetensors_available, logging
 from ...utils.import_utils import BACKENDS_MAPPING
 from ..latent_diffusion.pipeline_latent_diffusion import LDMBertConfig, LDMBertModel
 from ..paint_by_example import PaintByExampleImageEncoder
@@ -57,6 +58,10 @@ from .safety_checker import StableDiffusionSafetyChecker
 from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
 
 
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+    from accelerate.utils import set_module_tensor_to_device
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -770,11 +775,12 @@ def convert_ldm_bert_checkpoint(checkpoint, config):
 
 
 def convert_ldm_clip_checkpoint(checkpoint, local_files_only=False, text_encoder=None):
-    text_model = (
+    if text_encoder is None:
-        CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14", local_files_only=local_files_only)
+        config_name = "openai/clip-vit-large-patch14"
-        if text_encoder is None
+        config = CLIPTextConfig.from_pretrained(config_name)
-        else text_encoder
+
-    )
+        with init_empty_weights():
+            text_model = CLIPTextModel(config)
 
     keys = list(checkpoint.keys())
 
@@ -787,7 +793,8 @@ def convert_ldm_clip_checkpoint(checkpoint, local_files_only=False, text_encoder
             if key.startswith(prefix):
                 text_model_dict[key[len(prefix + ".") :]] = checkpoint[key]
 
-    text_model.load_state_dict(text_model_dict)
+    for param_name, param in text_model_dict.items():
+        set_module_tensor_to_device(text_model, param_name, "cpu", value=param)
 
     return text_model
 
@@ -884,14 +891,26 @@ def convert_paint_by_example_checkpoint(checkpoint):
     return model
 
 
-def convert_open_clip_checkpoint(checkpoint, prefix="cond_stage_model.model."):
+def convert_open_clip_checkpoint(
+    checkpoint, config_name, prefix="cond_stage_model.model.", has_projection=False, **config_kwargs
+):
     # text_model = CLIPTextModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="text_encoder")
-    text_model = CLIPTextModelWithProjection.from_pretrained(
+    # text_model = CLIPTextModelWithProjection.from_pretrained(
-        "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", projection_dim=1280
+    #    "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", projection_dim=1280
-    )
+    # )
+    config = CLIPTextConfig.from_pretrained(config_name, **config_kwargs)
+
+    with init_empty_weights():
+        text_model = CLIPTextModelWithProjection(config) if has_projection else CLIPTextModel(config)
 
     keys = list(checkpoint.keys())
 
+    keys_to_ignore = []
+    if config_name == "stabilityai/stable-diffusion-2" and config.num_hidden_layers == 23:
+        # make sure to remove all keys > 22
+        keys_to_ignore += [k for k in keys if k.startswith("cond_stage_model.model.transformer.resblocks.23")]
+        keys_to_ignore += ["cond_stage_model.model.text_projection"]
+
     text_model_dict = {}
 
     if prefix + "text_projection" in checkpoint:
@@ -902,8 +921,8 @@ def convert_open_clip_checkpoint(checkpoint, prefix="cond_stage_model.model."):
     text_model_dict["text_model.embeddings.position_ids"] = text_model.text_model.embeddings.get_buffer("position_ids")
 
     for key in keys:
-        # if "resblocks.23" in key:  # Diffusers drops the final layer and only uses the penultimate layer
+        if key in keys_to_ignore:
-        #     continue
+            continue
         if key[len(prefix) :] in textenc_conversion_map:
             if key.endswith("text_projection"):
                 value = checkpoint[key].T
@@ -931,7 +950,8 @@ def convert_open_clip_checkpoint(checkpoint, prefix="cond_stage_model.model."):
 
                 text_model_dict[new_key] = checkpoint[key]
 
-    text_model.load_state_dict(text_model_dict)
+    for param_name, param in text_model_dict.items():
+        set_module_tensor_to_device(text_model, param_name, "cpu", value=param)
 
     return text_model
 
@@ -1061,7 +1081,7 @@ def convert_controlnet_checkpoint(
 def download_from_original_stable_diffusion_ckpt(
     checkpoint_path: str,
     original_config_file: str = None,
-    image_size: int = 512,
+    image_size: Optional[int] = None,
     prediction_type: str = None,
     model_type: str = None,
     extract_ema: bool = False,
@@ -1144,6 +1164,7 @@ def download_from_original_stable_diffusion_ckpt(
         LDMTextToImagePipeline,
         PaintByExamplePipeline,
         StableDiffusionControlNetPipeline,
+        StableDiffusionInpaintPipeline,
         StableDiffusionPipeline,
         StableDiffusionXLImg2ImgPipeline,
         StableDiffusionXLPipeline,
@@ -1166,12 +1187,9 @@ def download_from_original_stable_diffusion_ckpt(
         if not is_safetensors_available():
             raise ValueError(BACKENDS_MAPPING["safetensors"][1])
 
-        from safetensors import safe_open
+        from safetensors.torch import load_file as safe_load
 
-        checkpoint = {}
+        checkpoint = safe_load(checkpoint_path, device="cpu")
-        with safe_open(checkpoint_path, framework="pt", device="cpu") as f:
-            for key in f.keys():
-                checkpoint[key] = f.get_tensor(key)
     else:
         if device is None:
             device = "cuda" if torch.cuda.is_available() else "cpu"
@@ -1183,7 +1201,7 @@ def download_from_original_stable_diffusion_ckpt(
     if "global_step" in checkpoint:
         global_step = checkpoint["global_step"]
     else:
-        logger.warning("global_step key not found in model")
+        logger.debug("global_step key not found in model")
         global_step = None
 
     # NOTE: this while loop isn't great but this controlnet checkpoint has one additional
@@ -1230,8 +1248,15 @@ def download_from_original_stable_diffusion_ckpt(
             model_type = "SDXL"
         else:
             model_type = "SDXL-Refiner"
+        if image_size is None:
+            image_size = 1024
 
-    if num_in_channels is not None:
+    if num_in_channels is None and pipeline_class == StableDiffusionInpaintPipeline:
+        num_in_channels = 9
+    elif num_in_channels is None:
+        num_in_channels = 4
+
+    if "unet_config" in original_config.model.params:
         original_config["model"]["params"]["unet_config"]["params"]["in_channels"] = num_in_channels
 
     if (
@@ -1263,7 +1288,6 @@ def download_from_original_stable_diffusion_ckpt(
     num_train_timesteps = getattr(original_config.model.params, "timesteps", None) or 1000
 
     if model_type in ["SDXL", "SDXL-Refiner"]:
-        image_size = 1024
         scheduler_dict = {
             "beta_schedule": "scaled_linear",
             "beta_start": 0.00085,
@@ -1279,7 +1303,6 @@ def download_from_original_stable_diffusion_ckpt(
         }
         scheduler = EulerDiscreteScheduler.from_config(scheduler_dict)
         scheduler_type = "euler"
-        vae_path = "stabilityai/sdxl-vae"
     else:
         beta_start = getattr(original_config.model.params, "linear_start", None) or 0.02
         beta_end = getattr(original_config.model.params, "linear_end", None) or 0.085
@@ -1318,25 +1341,45 @@ def download_from_original_stable_diffusion_ckpt(
     # Convert the UNet2DConditionModel model.
     unet_config = create_unet_diffusers_config(original_config, image_size=image_size)
     unet_config["upcast_attention"] = upcast_attention
-    unet = UNet2DConditionModel(**unet_config)
+    with init_empty_weights():
+        unet = UNet2DConditionModel(**unet_config)
 
     converted_unet_checkpoint = convert_ldm_unet_checkpoint(
         checkpoint, unet_config, path=checkpoint_path, extract_ema=extract_ema
     )
-    unet.load_state_dict(converted_unet_checkpoint)
+
+    for param_name, param in converted_unet_checkpoint.items():
+        set_module_tensor_to_device(unet, param_name, "cpu", value=param)
 
     # Convert the VAE model.
     if vae_path is None:
         vae_config = create_vae_diffusers_config(original_config, image_size=image_size)
         converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)
 
-        vae = AutoencoderKL(**vae_config)
+        if (
-        vae.load_state_dict(converted_vae_checkpoint)
+            "model" in original_config
+            and "params" in original_config.model
+            and "scale_factor" in original_config.model.params
+        ):
+            vae_scaling_factor = original_config.model.params.scale_factor
+        else:
+            vae_scaling_factor = 0.18215  # default SD scaling factor
+
+        vae_config["scaling_factor"] = vae_scaling_factor
+
+        with init_empty_weights():
+            vae = AutoencoderKL(**vae_config)
+
+        for param_name, param in converted_vae_checkpoint.items():
+            set_module_tensor_to_device(vae, param_name, "cpu", value=param)
     else:
         vae = AutoencoderKL.from_pretrained(vae_path)
 
     if model_type == "FrozenOpenCLIPEmbedder":
-        text_model = convert_open_clip_checkpoint(checkpoint)
+        config_name = "stabilityai/stable-diffusion-2"
+        config_kwargs = {"subfolder": "text_encoder"}
+
+        text_model = convert_open_clip_checkpoint(checkpoint, config_name, **config_kwargs)
         tokenizer = CLIPTokenizer.from_pretrained("stabilityai/stable-diffusion-2", subfolder="tokenizer")
 
         if stable_unclip is None:
@@ -1469,7 +1512,12 @@ def download_from_original_stable_diffusion_ckpt(
             tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
             text_encoder = convert_ldm_clip_checkpoint(checkpoint, local_files_only=local_files_only)
             tokenizer_2 = CLIPTokenizer.from_pretrained("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", pad_token="!")
-            text_encoder_2 = convert_open_clip_checkpoint(checkpoint, prefix="conditioner.embedders.1.model.")
+
+            config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
+            config_kwargs = {"projection_dim": 1280}
+            text_encoder_2 = convert_open_clip_checkpoint(
+                checkpoint, config_name, prefix="conditioner.embedders.1.model.", has_projection=True, **config_kwargs
+            )
 
             pipe = StableDiffusionXLPipeline(
                 vae=vae,
@@ -1485,7 +1533,12 @@ def download_from_original_stable_diffusion_ckpt(
             tokenizer = None
             text_encoder = None
             tokenizer_2 = CLIPTokenizer.from_pretrained("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", pad_token="!")
-            text_encoder_2 = convert_open_clip_checkpoint(checkpoint, prefix="conditioner.embedders.0.model.")
+
+            config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
+            config_kwargs = {"projection_dim": 1280}
+            text_encoder_2 = convert_open_clip_checkpoint(
+                checkpoint, config_name, prefix="conditioner.embedders.0.model.", has_projection=True, **config_kwargs
+            )
 
             pipe = StableDiffusionXLImg2ImgPipeline(
                 vae=vae,

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -24,7 +24,7 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from ...configuration_utils import FrozenDict
 from ...image_processor import VaeImageProcessor
-from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import deprecate, is_accelerate_available, is_accelerate_version, logging, randn_tensor
@@ -153,7 +153,9 @@ def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool
     return mask, masked_image
 
 
-class StableDiffusionInpaintPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
+class StableDiffusionInpaintPipeline(
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+):
     r"""
     Pipeline for text-guided image inpainting using Stable Diffusion.
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py

@@ -748,15 +748,19 @@ class StableDiffusionUpscalePipeline(DiffusionPipeline, TextualInversionLoaderMi
         # make sure the VAE is in float32 mode, as it overflows in float16
         self.vae.to(dtype=torch.float32)
 
-        use_torch_2_0_or_xformers = self.vae.decoder.mid_block.attentions[0].processor in [
+        use_torch_2_0_or_xformers = isinstance(
-            AttnProcessor2_0,
+            self.vae.decoder.mid_block.attentions[0].processor,
-            XFormersAttnProcessor,
+            (
-            LoRAXFormersAttnProcessor,
+                AttnProcessor2_0,
-            LoRAAttnProcessor2_0,
+                XFormersAttnProcessor,
-        ]
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+            ),
+        )
+
         # if xformers or torch_2_0 is used attention block does not need
         # to be in float32 which can save lots of memory
-        if not use_torch_2_0_or_xformers:
+        if use_torch_2_0_or_xformers:
             self.vae.post_quant_conv.to(latents.dtype)
             self.vae.decoder.conv_in.to(latents.dtype)
             self.vae.decoder.mid_block.to(latents.dtype)

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py

@@ -129,6 +129,7 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
         self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.default_sample_size = self.unet.config.sample_size
 
         self.watermark = StableDiffusionXLWatermarker()
 
@@ -652,8 +653,8 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
             "not-safe-for-work" (nsfw) content, according to the `safety_checker`.
         """
         # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        height = height or self.default_sample_size * self.vae_scale_factor
-        width = width or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
 
         original_size = original_size or (height, width)
         target_size = target_size or (height, width)
@@ -785,15 +786,18 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin):
         # make sure the VAE is in float32 mode, as it overflows in float16
         self.vae.to(dtype=torch.float32)
 
-        use_torch_2_0_or_xformers = self.vae.decoder.mid_block.attentions[0].processor in [
+        use_torch_2_0_or_xformers = isinstance(
-            AttnProcessor2_0,
+            self.vae.decoder.mid_block.attentions[0].processor,
-            XFormersAttnProcessor,
+            (
-            LoRAXFormersAttnProcessor,
+                AttnProcessor2_0,
-            LoRAAttnProcessor2_0,
+                XFormersAttnProcessor,
-        ]
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+            ),
+        )
         # if xformers or torch_2_0 is used attention block does not need
         # to be in float32 which can save lots of memory
-        if not use_torch_2_0_or_xformers:
+        if use_torch_2_0_or_xformers:
             self.vae.post_quant_conv.to(latents.dtype)
             self.vae.decoder.conv_in.to(latents.dtype)
             self.vae.decoder.mid_block.to(latents.dtype)

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

@@ -859,15 +859,18 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin):
         # make sure the VAE is in float32 mode, as it overflows in float16
         self.vae.to(dtype=torch.float32)
 
-        use_torch_2_0_or_xformers = self.vae.decoder.mid_block.attentions[0].processor in [
+        use_torch_2_0_or_xformers = isinstance(
-            AttnProcessor2_0,
+            self.vae.decoder.mid_block.attentions[0].processor,
-            XFormersAttnProcessor,
+            (
-            LoRAXFormersAttnProcessor,
+                AttnProcessor2_0,
-            LoRAAttnProcessor2_0,
+                XFormersAttnProcessor,
-        ]
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+            ),
+        )
         # if xformers or torch_2_0 is used attention block does not need
         # to be in float32 which can save lots of memory
-        if not use_torch_2_0_or_xformers:
+        if use_torch_2_0_or_xformers:
             self.vae.post_quant_conv.to(latents.dtype)
             self.vae.decoder.conv_in.to(latents.dtype)
             self.vae.decoder.mid_block.to(latents.dtype)

tests/others/test_config.py

@@ -264,7 +264,7 @@ class ConfigTester(unittest.TestCase):
         config_dict = {k: v for k, v in config.config.items() if not k.startswith("_")}
 
         # make sure that default config has all keys in `_use_default_values`
-        assert set(config_dict.keys()) == config.config._use_default_values
+        assert set(config_dict.keys()) == set(config.config._use_default_values)
 
         with tempfile.TemporaryDirectory() as tmpdirname:
             config.save_config(tmpdirname)

tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint.py

@@ -20,17 +20,20 @@ import unittest
 
 import numpy as np
 import torch
+from huggingface_hub import hf_hub_download
 from PIL import Image
 from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
 
 from diffusers import (
     AutoencoderKL,
+    DDIMScheduler,
     DPMSolverMultistepScheduler,
     LMSDiscreteScheduler,
     PNDMScheduler,
     StableDiffusionInpaintPipeline,
     UNet2DConditionModel,
 )
+from diffusers.models.attention_processor import AttnProcessor
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint import prepare_mask_and_masked_image
 from diffusers.utils import floats_tensor, load_image, load_numpy, nightly, slow, torch_device
 from diffusers.utils.testing_utils import (
@@ -512,6 +515,42 @@ class StableDiffusionInpaintPipelineSlowTests(unittest.TestCase):
 
         assert np.abs(expected_slice - image_slice).max() < 6e-4
 
+    def test_download_local(self):
+        filename = hf_hub_download("runwayml/stable-diffusion-inpainting", filename="sd-v1-5-inpainting.ckpt")
+
+        pipe = StableDiffusionInpaintPipeline.from_single_file(filename, torch_dtype=torch.float16)
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.to("cuda")
+
+        inputs = self.get_inputs(torch_device)
+        inputs["num_inference_steps"] = 1
+        image_out = pipe(**inputs).images[0]
+
+        assert image_out.shape == (512, 512, 3)
+
+    def test_download_ckpt_diff_format_is_same(self):
+        ckpt_path = "https://huggingface.co/runwayml/stable-diffusion-inpainting/blob/main/sd-v1-5-inpainting.ckpt"
+
+        pipe = StableDiffusionInpaintPipeline.from_single_file(ckpt_path)
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.unet.set_attn_processor(AttnProcessor())
+        pipe.to("cuda")
+
+        inputs = self.get_inputs(torch_device)
+        inputs["num_inference_steps"] = 5
+        image_ckpt = pipe(**inputs).images[0]
+
+        pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.unet.set_attn_processor(AttnProcessor())
+        pipe.to("cuda")
+
+        inputs = self.get_inputs(torch_device)
+        inputs["num_inference_steps"] = 5
+        image = pipe(**inputs).images[0]
+
+        assert np.max(np.abs(image - image_ckpt)) < 1e-4
+
 
 @nightly
 @require_torch_gpu

tests/pipelines/stable_diffusion_2/test_stable_diffusion_v_pred.py

@@ -19,6 +19,7 @@ import unittest
 
 import numpy as np
 import torch
+from huggingface_hub import hf_hub_download
 from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
 
 from diffusers import (
@@ -29,6 +30,7 @@ from diffusers import (
     StableDiffusionPipeline,
     UNet2DConditionModel,
 )
+from diffusers.models.attention_processor import AttnProcessor
 from diffusers.utils import load_numpy, slow, torch_device
 from diffusers.utils.testing_utils import enable_full_determinism, require_torch_gpu
 
@@ -426,6 +428,40 @@ class StableDiffusion2VPredictionPipelineIntegrationTests(unittest.TestCase):
         assert image.shape == (768, 768, 3)
         assert np.abs(expected_image - image).max() < 7.5e-1
 
+    def test_download_local(self):
+        filename = hf_hub_download("stabilityai/stable-diffusion-2-1", filename="v2-1_768-ema-pruned.safetensors")
+
+        pipe = StableDiffusionPipeline.from_single_file(filename, torch_dtype=torch.float16)
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.to("cuda")
+
+        image_out = pipe("test", num_inference_steps=1, output_type="np").images[0]
+
+        assert image_out.shape == (768, 768, 3)
+
+    def test_download_ckpt_diff_format_is_same(self):
+        single_file_path = (
+            "https://huggingface.co/stabilityai/stable-diffusion-2-1/blob/main/v2-1_768-ema-pruned.safetensors"
+        )
+
+        pipe_single = StableDiffusionPipeline.from_single_file(single_file_path)
+        pipe_single.scheduler = DDIMScheduler.from_config(pipe_single.scheduler.config)
+        pipe_single.unet.set_attn_processor(AttnProcessor())
+        pipe_single.to("cuda")
+
+        generator = torch.Generator(device="cpu").manual_seed(0)
+        image_ckpt = pipe_single("a turtle", num_inference_steps=5, generator=generator, output_type="np").images[0]
+
+        pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1")
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.unet.set_attn_processor(AttnProcessor())
+        pipe.to("cuda")
+
+        generator = torch.Generator(device="cpu").manual_seed(0)
+        image = pipe("a turtle", num_inference_steps=5, generator=generator, output_type="np").images[0]
+
+        assert np.max(np.abs(image - image_ckpt)) < 1e-3
+
     def test_stable_diffusion_text2img_intermediate_state_v_pred(self):
         number_of_steps = 0
 

tests/pipelines/test_pipelines.py

@@ -14,6 +14,7 @@
 # limitations under the License.
 
 import gc
+import glob
 import json
 import os
 import random
@@ -56,6 +57,7 @@ from diffusers import (
     UniPCMultistepScheduler,
     logging,
 )
+from diffusers.pipelines.pipeline_utils import variant_compatible_siblings
 from diffusers.schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
 from diffusers.utils import (
     CONFIG_NAME,
@@ -1361,6 +1363,29 @@ class PipelineFastTests(unittest.TestCase):
             assert sd.config.safety_checker != (None, None)
             assert sd.config.feature_extractor != (None, None)
 
+    def test_warning_no_variant_available(self):
+        variant = "fp16"
+        with self.assertWarns(FutureWarning) as warning_context:
+            cached_folder = StableDiffusionPipeline.download(
+                "hf-internal-testing/diffusers-stable-diffusion-tiny-all", variant=variant
+            )
+
+        assert "but no such modeling files are available" in str(warning_context.warning)
+        assert variant in str(warning_context.warning)
+
+        def get_all_filenames(directory):
+            filenames = glob.glob(directory + "/**", recursive=True)
+            filenames = [f for f in filenames if os.path.isfile(f)]
+            return filenames
+
+        filenames = get_all_filenames(str(cached_folder))
+
+        all_model_files, variant_model_files = variant_compatible_siblings(filenames, variant=variant)
+
+        # make sure that none of the model names are variant model names
+        assert len(variant_model_files) == 0
+        assert len(all_model_files) > 0
+
 
 @slow
 @require_torch_gpu