diffusers 0.24.0__py3-none-any.whl → 0.25.1__py3-none-any.whl

diffusers/loaders/single_file.py

@@ -18,10 +18,9 @@ from pathlib import Path
 import requests
 import torch
 from huggingface_hub import hf_hub_download
+from huggingface_hub.utils import validate_hf_hub_args
 
 from ..utils import (
-    DIFFUSERS_CACHE,
-    HF_HUB_OFFLINE,
     deprecate,
     is_accelerate_available,
     is_omegaconf_available,
@@ -52,6 +51,7 @@ class FromSingleFileMixin:
         return cls.from_single_file(*args, **kwargs)
 
     @classmethod
+    @validate_hf_hub_args
     def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
         r"""
         Instantiate a [`DiffusionPipeline`] from pretrained pipeline weights saved in the `.ckpt` or `.safetensors`
@@ -81,7 +81,7 @@ class FromSingleFileMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -154,12 +154,12 @@ class FromSingleFileMixin:
 
         original_config_file = kwargs.pop("original_config_file", None)
         config_files = kwargs.pop("config_files", None)
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         extract_ema = kwargs.pop("extract_ema", False)
         image_size = kwargs.pop("image_size", None)
@@ -169,10 +169,12 @@ class FromSingleFileMixin:
         load_safety_checker = kwargs.pop("load_safety_checker", True)
         prediction_type = kwargs.pop("prediction_type", None)
         text_encoder = kwargs.pop("text_encoder", None)
+        text_encoder_2 = kwargs.pop("text_encoder_2", None)
         vae = kwargs.pop("vae", None)
         controlnet = kwargs.pop("controlnet", None)
         adapter = kwargs.pop("adapter", None)
         tokenizer = kwargs.pop("tokenizer", None)
+        tokenizer_2 = kwargs.pop("tokenizer_2", None)
 
         torch_dtype = kwargs.pop("torch_dtype", None)
 
@@ -253,7 +255,7 @@ class FromSingleFileMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 force_download=force_download,
             )
@@ -274,15 +276,17 @@ class FromSingleFileMixin:
             load_safety_checker=load_safety_checker,
             prediction_type=prediction_type,
             text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
             vae=vae,
             tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
             original_config_file=original_config_file,
             config_files=config_files,
             local_files_only=local_files_only,
         )
 
         if torch_dtype is not None:
-            pipe.to(torch_dtype=torch_dtype)
+            pipe.to(dtype=torch_dtype)
 
         return pipe
 
@@ -293,6 +297,7 @@ class FromOriginalVAEMixin:
     """
 
     @classmethod
+    @validate_hf_hub_args
     def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
         r"""
         Instantiate a [`AutoencoderKL`] from pretrained ControlNet weights saved in the original `.ckpt` or
@@ -322,7 +327,7 @@ class FromOriginalVAEMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to True, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -379,12 +384,12 @@ class FromOriginalVAEMixin:
         )
 
         config_file = kwargs.pop("config_file", None)
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         image_size = kwargs.pop("image_size", None)
         scaling_factor = kwargs.pop("scaling_factor", None)
@@ -425,7 +430,7 @@ class FromOriginalVAEMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 force_download=force_download,
             )
@@ -490,6 +495,7 @@ class FromOriginalControlnetMixin:
     """
 
     @classmethod
+    @validate_hf_hub_args
     def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
         r"""
         Instantiate a [`ControlNetModel`] from pretrained ControlNet weights saved in the original `.ckpt` or
@@ -519,7 +525,7 @@ class FromOriginalControlnetMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to True, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -555,12 +561,12 @@ class FromOriginalControlnetMixin:
         from ..pipelines.stable_diffusion.convert_from_ckpt import download_controlnet_from_original_ckpt
 
         config_file = kwargs.pop("config_file", None)
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         num_in_channels = kwargs.pop("num_in_channels", None)
         use_linear_projection = kwargs.pop("use_linear_projection", None)
         revision = kwargs.pop("revision", None)
@@ -603,7 +609,7 @@ class FromOriginalControlnetMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
+                token=token,
                 revision=revision,
                 force_download=force_download,
             )

diffusers/loaders/textual_inversion.py

@@ -15,16 +15,10 @@ from typing import Dict, List, Optional, Union
 
 import safetensors
 import torch
+from huggingface_hub.utils import validate_hf_hub_args
 from torch import nn
 
-from ..utils import (
-    DIFFUSERS_CACHE,
-    HF_HUB_OFFLINE,
-    _get_model_file,
-    is_accelerate_available,
-    is_transformers_available,
-    logging,
-)
+from ..utils import _get_model_file, is_accelerate_available, is_transformers_available, logging
 
 
 if is_transformers_available():
@@ -39,13 +33,14 @@ TEXT_INVERSION_NAME = "learned_embeds.bin"
 TEXT_INVERSION_NAME_SAFE = "learned_embeds.safetensors"
 
 
+@validate_hf_hub_args
 def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs):
-    cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+    cache_dir = kwargs.pop("cache_dir", None)
     force_download = kwargs.pop("force_download", False)
     resume_download = kwargs.pop("resume_download", False)
     proxies = kwargs.pop("proxies", None)
-    local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-    use_auth_token = kwargs.pop("use_auth_token", None)
+    local_files_only = kwargs.pop("local_files_only", None)
+    token = kwargs.pop("token", None)
     revision = kwargs.pop("revision", None)
     subfolder = kwargs.pop("subfolder", None)
     weight_name = kwargs.pop("weight_name", None)
@@ -79,7 +74,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                         resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        use_auth_token=use_auth_token,
+                        token=token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -100,7 +95,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
+                    token=token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,
@@ -267,6 +262,7 @@ class TextualInversionLoaderMixin:
 
         return all_tokens, all_embeddings
 
+    @validate_hf_hub_args
     def load_textual_inversion(
         self,
         pretrained_model_name_or_path: Union[str, List[str], Dict[str, torch.Tensor], List[Dict[str, torch.Tensor]]],
@@ -320,7 +316,7 @@ class TextualInversionLoaderMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):

diffusers/loaders/unet.py

@@ -11,21 +11,22 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import inspect
 import os
 from collections import defaultdict
 from contextlib import nullcontext
+from functools import partial
 from typing import Callable, Dict, List, Optional, Union
 
 import safetensors
 import torch
 import torch.nn.functional as F
+from huggingface_hub.utils import validate_hf_hub_args
 from torch import nn
 
-from ..models.embeddings import ImageProjection
+from ..models.embeddings import ImageProjection, MLPProjection, Resampler
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
 from ..utils import (
-    DIFFUSERS_CACHE,
-    HF_HUB_OFFLINE,
     USE_PEFT_BACKEND,
     _get_model_file,
     delete_adapter_layers,
@@ -62,6 +63,7 @@ class UNet2DConditionLoadersMixin:
     text_encoder_name = TEXT_ENCODER_NAME
     unet_name = UNET_NAME
 
+    @validate_hf_hub_args
     def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
         r"""
         Load pretrained attention processor layers into [`UNet2DConditionModel`]. Attention processor layers have to be
@@ -95,7 +97,7 @@ class UNet2DConditionLoadersMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            use_auth_token (`str` or *bool*, *optional*):
+            token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
@@ -130,12 +132,12 @@ class UNet2DConditionLoadersMixin:
         from ..models.attention_processor import CustomDiffusionAttnProcessor
         from ..models.lora import LoRACompatibleConv, LoRACompatibleLinear, LoRAConv2dLayer, LoRALinearLayer
 
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
-        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         subfolder = kwargs.pop("subfolder", None)
         weight_name = kwargs.pop("weight_name", None)
@@ -184,7 +186,7 @@ class UNet2DConditionLoadersMixin:
                         resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        use_auth_token=use_auth_token,
+                        token=token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -204,7 +206,7 @@ class UNet2DConditionLoadersMixin:
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
+                    token=token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,
@@ -504,22 +506,43 @@ class UNet2DConditionLoadersMixin:
         save_function(state_dict, os.path.join(save_directory, weight_name))
         logger.info(f"Model weights saved in {os.path.join(save_directory, weight_name)}")
 
-    def fuse_lora(self, lora_scale=1.0, safe_fusing=False):
+    def fuse_lora(self, lora_scale=1.0, safe_fusing=False, adapter_names=None):
         self.lora_scale = lora_scale
         self._safe_fusing = safe_fusing
-        self.apply(self._fuse_lora_apply)
+        self.apply(partial(self._fuse_lora_apply, adapter_names=adapter_names))
 
-    def _fuse_lora_apply(self, module):
+    def _fuse_lora_apply(self, module, adapter_names=None):
         if not USE_PEFT_BACKEND:
             if hasattr(module, "_fuse_lora"):
                 module._fuse_lora(self.lora_scale, self._safe_fusing)
+
+            if adapter_names is not None:
+                raise ValueError(
+                    "The `adapter_names` argument is not supported in your environment. Please switch"
+                    " to PEFT backend to use this argument by installing latest PEFT and transformers."
+                    " `pip install -U peft transformers`"
+                )
         else:
             from peft.tuners.tuners_utils import BaseTunerLayer
 
+            merge_kwargs = {"safe_merge": self._safe_fusing}
+
             if isinstance(module, BaseTunerLayer):
                 if self.lora_scale != 1.0:
                     module.scale_layer(self.lora_scale)
-                module.merge(safe_merge=self._safe_fusing)
+
+                # For BC with prevous PEFT versions, we need to check the signature
+                # of the `merge` method to see if it supports the `adapter_names` argument.
+                supported_merge_kwargs = list(inspect.signature(module.merge).parameters)
+                if "adapter_names" in supported_merge_kwargs:
+                    merge_kwargs["adapter_names"] = adapter_names
+                elif "adapter_names" not in supported_merge_kwargs and adapter_names is not None:
+                    raise ValueError(
+                        "The `adapter_names` argument is not supported with your PEFT version. Please upgrade"
+                        " to the latest version of PEFT. `pip install -U peft`"
+                    )
+
+                module.merge(**merge_kwargs)
 
     def unfuse_lora(self):
         self.apply(self._unfuse_lora_apply)
@@ -664,6 +687,80 @@ class UNet2DConditionLoadersMixin:
             if hasattr(self, "peft_config"):
                 self.peft_config.pop(adapter_name, None)
 
+    def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict):
+        updated_state_dict = {}
+        image_projection = None
+
+        if "proj.weight" in state_dict:
+            # IP-Adapter
+            num_image_text_embeds = 4
+            clip_embeddings_dim = state_dict["proj.weight"].shape[-1]
+            cross_attention_dim = state_dict["proj.weight"].shape[0] // 4
+
+            image_projection = ImageProjection(
+                cross_attention_dim=cross_attention_dim,
+                image_embed_dim=clip_embeddings_dim,
+                num_image_text_embeds=num_image_text_embeds,
+            )
+
+            for key, value in state_dict.items():
+                diffusers_name = key.replace("proj", "image_embeds")
+                updated_state_dict[diffusers_name] = value
+
+        elif "proj.3.weight" in state_dict:
+            # IP-Adapter Full
+            clip_embeddings_dim = state_dict["proj.0.weight"].shape[0]
+            cross_attention_dim = state_dict["proj.3.weight"].shape[0]
+
+            image_projection = MLPProjection(
+                cross_attention_dim=cross_attention_dim, image_embed_dim=clip_embeddings_dim
+            )
+
+            for key, value in state_dict.items():
+                diffusers_name = key.replace("proj.0", "ff.net.0.proj")
+                diffusers_name = diffusers_name.replace("proj.2", "ff.net.2")
+                diffusers_name = diffusers_name.replace("proj.3", "norm")
+                updated_state_dict[diffusers_name] = value
+
+        else:
+            # IP-Adapter Plus
+            num_image_text_embeds = state_dict["latents"].shape[1]
+            embed_dims = state_dict["proj_in.weight"].shape[1]
+            output_dims = state_dict["proj_out.weight"].shape[0]
+            hidden_dims = state_dict["latents"].shape[2]
+            heads = state_dict["layers.0.0.to_q.weight"].shape[0] // 64
+
+            image_projection = Resampler(
+                embed_dims=embed_dims,
+                output_dims=output_dims,
+                hidden_dims=hidden_dims,
+                heads=heads,
+                num_queries=num_image_text_embeds,
+            )
+
+            for key, value in state_dict.items():
+                diffusers_name = key.replace("0.to", "2.to")
+                diffusers_name = diffusers_name.replace("1.0.weight", "3.0.weight")
+                diffusers_name = diffusers_name.replace("1.0.bias", "3.0.bias")
+                diffusers_name = diffusers_name.replace("1.1.weight", "3.1.net.0.proj.weight")
+                diffusers_name = diffusers_name.replace("1.3.weight", "3.1.net.2.weight")
+
+                if "norm1" in diffusers_name:
+                    updated_state_dict[diffusers_name.replace("0.norm1", "0")] = value
+                elif "norm2" in diffusers_name:
+                    updated_state_dict[diffusers_name.replace("0.norm2", "1")] = value
+                elif "to_kv" in diffusers_name:
+                    v_chunk = value.chunk(2, dim=0)
+                    updated_state_dict[diffusers_name.replace("to_kv", "to_k")] = v_chunk[0]
+                    updated_state_dict[diffusers_name.replace("to_kv", "to_v")] = v_chunk[1]
+                elif "to_out" in diffusers_name:
+                    updated_state_dict[diffusers_name.replace("to_out", "to_out.0")] = value
+                else:
+                    updated_state_dict[diffusers_name] = value
+
+        image_projection.load_state_dict(updated_state_dict)
+        return image_projection
+
     def _load_ip_adapter_weights(self, state_dict):
         from ..models.attention_processor import (
             AttnProcessor,
@@ -672,6 +769,20 @@ class UNet2DConditionLoadersMixin:
             IPAdapterAttnProcessor2_0,
         )
 
+        if "proj.weight" in state_dict["image_proj"]:
+            # IP-Adapter
+            num_image_text_embeds = 4
+        elif "proj.3.weight" in state_dict["image_proj"]:
+            # IP-Adapter Full Face
+            num_image_text_embeds = 257  # 256 CLIP tokens + 1 CLS token
+        else:
+            # IP-Adapter Plus
+            num_image_text_embeds = state_dict["image_proj"]["latents"].shape[1]
+
+        # Set encoder_hid_proj after loading ip_adapter weights,
+        # because `Resampler` also has `attn_processors`.
+        self.encoder_hid_proj = None
+
         # set ip-adapter cross-attention processors & load state_dict
         attn_procs = {}
         key_id = 1
@@ -695,7 +806,10 @@ class UNet2DConditionLoadersMixin:
                     IPAdapterAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else IPAdapterAttnProcessor
                 )
                 attn_procs[name] = attn_processor_class(
-                    hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, scale=1.0
+                    hidden_size=hidden_size,
+                    cross_attention_dim=cross_attention_dim,
+                    scale=1.0,
+                    num_tokens=num_image_text_embeds,
                 ).to(dtype=self.dtype, device=self.device)
 
                 value_dict = {}
@@ -707,29 +821,8 @@ class UNet2DConditionLoadersMixin:
 
         self.set_attn_processor(attn_procs)
 
-        # create image projection layers.
-        clip_embeddings_dim = state_dict["image_proj"]["proj.weight"].shape[-1]
-        cross_attention_dim = state_dict["image_proj"]["proj.weight"].shape[0] // 4
-
-        image_projection = ImageProjection(
-            cross_attention_dim=cross_attention_dim, image_embed_dim=clip_embeddings_dim, num_image_text_embeds=4
-        )
-        image_projection.to(dtype=self.dtype, device=self.device)
-
-        # load image projection layer weights
-        image_proj_state_dict = {}
-        image_proj_state_dict.update(
-            {
-                "image_embeds.weight": state_dict["image_proj"]["proj.weight"],
-                "image_embeds.bias": state_dict["image_proj"]["proj.bias"],
-                "norm.weight": state_dict["image_proj"]["norm.weight"],
-                "norm.bias": state_dict["image_proj"]["norm.bias"],
-            }
-        )
-
-        image_projection.load_state_dict(image_proj_state_dict)
+        # convert IP-Adapter Image Projection layers to diffusers
+        image_projection = self._convert_ip_adapter_image_proj_to_diffusers(state_dict["image_proj"])
 
         self.encoder_hid_proj = image_projection.to(device=self.device, dtype=self.dtype)
         self.config.encoder_hid_dim_type = "ip_image_proj"
-
-    delete_adapter_layers

diffusers/models/__init__.py

@@ -26,13 +26,14 @@ _import_structure = {}
 
 if is_torch_available():
     _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
-    _import_structure["autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
-    _import_structure["autoencoder_kl"] = ["AutoencoderKL"]
-    _import_structure["autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
-    _import_structure["autoencoder_tiny"] = ["AutoencoderTiny"]
-    _import_structure["consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
+    _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
+    _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
+    _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["controlnet"] = ["ControlNetModel"]
     _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
+    _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
     _import_structure["prior_transformer"] = ["PriorTransformer"]
     _import_structure["t5_film_transformer"] = ["T5FilmDecoder"]
@@ -42,9 +43,10 @@ if is_torch_available():
     _import_structure["unet_2d"] = ["UNet2DModel"]
     _import_structure["unet_2d_condition"] = ["UNet2DConditionModel"]
     _import_structure["unet_3d_condition"] = ["UNet3DConditionModel"]
-    _import_structure["unet_kandi3"] = ["Kandinsky3UNet"]
+    _import_structure["unet_kandinsky3"] = ["Kandinsky3UNet"]
     _import_structure["unet_motion_model"] = ["MotionAdapter", "UNetMotionModel"]
     _import_structure["unet_spatio_temporal_condition"] = ["UNetSpatioTemporalConditionModel"]
+    _import_structure["uvit_2d"] = ["UVit2DModel"]
     _import_structure["vq_model"] = ["VQModel"]
 
 if is_flax_available():
@@ -56,13 +58,16 @@ if is_flax_available():
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     if is_torch_available():
         from .adapter import MultiAdapter, T2IAdapter
-        from .autoencoder_asym_kl import AsymmetricAutoencoderKL
-        from .autoencoder_kl import AutoencoderKL
-        from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
-        from .autoencoder_tiny import AutoencoderTiny
-        from .consistency_decoder_vae import ConsistencyDecoderVAE
+        from .autoencoders import (
+            AsymmetricAutoencoderKL,
+            AutoencoderKL,
+            AutoencoderKLTemporalDecoder,
+            AutoencoderTiny,
+            ConsistencyDecoderVAE,
+        )
         from .controlnet import ControlNetModel
         from .dual_transformer_2d import DualTransformer2DModel
+        from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin
         from .prior_transformer import PriorTransformer
         from .t5_film_transformer import T5FilmDecoder
@@ -72,9 +77,10 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .unet_2d import UNet2DModel
         from .unet_2d_condition import UNet2DConditionModel
         from .unet_3d_condition import UNet3DConditionModel
-        from .unet_kandi3 import Kandinsky3UNet
+        from .unet_kandinsky3 import Kandinsky3UNet
         from .unet_motion_model import MotionAdapter, UNetMotionModel
         from .unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel
+        from .uvit_2d import UVit2DModel
         from .vq_model import VQModel
 
     if is_flax_available():

diffusers/models/activations.py

@@ -55,11 +55,12 @@ class GELU(nn.Module):
         dim_in (`int`): The number of channels in the input.
         dim_out (`int`): The number of channels in the output.
         approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
     """
 
-    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none"):
+    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True):
         super().__init__()
-        self.proj = nn.Linear(dim_in, dim_out)
+        self.proj = nn.Linear(dim_in, dim_out, bias=bias)
         self.approximate = approximate
 
     def gelu(self, gate: torch.Tensor) -> torch.Tensor:
@@ -81,13 +82,14 @@ class GEGLU(nn.Module):
     Parameters:
         dim_in (`int`): The number of channels in the input.
         dim_out (`int`): The number of channels in the output.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
     """
 
-    def __init__(self, dim_in: int, dim_out: int):
+    def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
         super().__init__()
         linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
 
-        self.proj = linear_cls(dim_in, dim_out * 2)
+        self.proj = linear_cls(dim_in, dim_out * 2, bias=bias)
 
     def gelu(self, gate: torch.Tensor) -> torch.Tensor:
         if gate.device.type != "mps":
@@ -109,11 +111,12 @@ class ApproximateGELU(nn.Module):
     Parameters:
         dim_in (`int`): The number of channels in the input.
         dim_out (`int`): The number of channels in the output.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
     """
 
-    def __init__(self, dim_in: int, dim_out: int):
+    def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
         super().__init__()
-        self.proj = nn.Linear(dim_in, dim_out)
+        self.proj = nn.Linear(dim_in, dim_out, bias=bias)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         x = self.proj(x)