diffusers 0.33.1__py3-none-any.whl → 0.34.0__py3-none-any.whl

diffusers/models/controlnets/controlnet_flux.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved.
+# Copyright 2025 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -20,12 +20,12 @@ import torch.nn as nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import PeftAdapterMixin
-from ...models.attention_processor import AttentionProcessor
-from ...models.modeling_utils import ModelMixin
 from ...utils import USE_PEFT_BACKEND, BaseOutput, logging, scale_lora_layers, unscale_lora_layers
+from ..attention_processor import AttentionProcessor
 from ..controlnets.controlnet import ControlNetConditioningEmbedding, zero_module
 from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
 from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
 from ..transformers.transformer_flux import FluxSingleTransformerBlock, FluxTransformerBlock
 
 
@@ -430,7 +430,7 @@ class FluxMultiControlNetModel(ModelMixin):
     ) -> Union[FluxControlNetOutput, Tuple]:
         # ControlNet-Union with multiple conditions
         # only load one ControlNet for saving memories
-        if len(self.nets) == 1 and self.nets[0].union:
+        if len(self.nets) == 1:
             controlnet = self.nets[0]
 
             for i, (image, mode, scale) in enumerate(zip(controlnet_cond, controlnet_mode, conditioning_scale)):
@@ -454,17 +454,18 @@ class FluxMultiControlNetModel(ModelMixin):
                     control_block_samples = block_samples
                     control_single_block_samples = single_block_samples
                 else:
-                    control_block_samples = [
-                        control_block_sample + block_sample
-                        for control_block_sample, block_sample in zip(control_block_samples, block_samples)
-                    ]
-
-                    control_single_block_samples = [
-                        control_single_block_sample + block_sample
-                        for control_single_block_sample, block_sample in zip(
-                            control_single_block_samples, single_block_samples
-                        )
-                    ]
+                    if block_samples is not None and control_block_samples is not None:
+                        control_block_samples = [
+                            control_block_sample + block_sample
+                            for control_block_sample, block_sample in zip(control_block_samples, block_samples)
+                        ]
+                    if single_block_samples is not None and control_single_block_samples is not None:
+                        control_single_block_samples = [
+                            control_single_block_sample + block_sample
+                            for control_single_block_sample, block_sample in zip(
+                                control_single_block_samples, single_block_samples
+                            )
+                        ]
 
         # Regular Multi-ControlNets
         # load all ControlNets into memories

diffusers/models/controlnets/controlnet_hunyuan.py

@@ -1,4 +1,4 @@
-# Copyright 2024 HunyuanDiT Authors, Qixun Wang and The HuggingFace Team. All rights reserved.
+# Copyright 2025 HunyuanDiT Authors, Qixun Wang and The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -103,7 +103,7 @@ class HunyuanDiT2DControlNetModel(ModelMixin, ConfigMixin):
                     activation_fn=activation_fn,
                     ff_inner_dim=int(self.inner_dim * mlp_ratio),
                     cross_attention_dim=cross_attention_dim,
-                    qk_norm=True,  # See http://arxiv.org/abs/2302.05442 for details.
+                    qk_norm=True,  # See https://huggingface.co/papers/2302.05442 for details.
                     skip=False,  # always False as it is the first half of the model
                 )
                 for layer in range(transformer_num_layers // 2 - 1)

diffusers/models/controlnets/controlnet_sana.py

@@ -0,0 +1,290 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, BaseOutput, logging, scale_lora_layers, unscale_lora_layers
+from ..attention_processor import AttentionProcessor
+from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormSingle, RMSNorm
+from ..transformers.sana_transformer import SanaTransformerBlock
+from .controlnet import zero_module
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class SanaControlNetOutput(BaseOutput):
+    controlnet_block_samples: Tuple[torch.Tensor]
+
+
+class SanaControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["SanaTransformerBlock", "PatchEmbed"]
+    _skip_layerwise_casting_patterns = ["patch_embed", "norm"]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 32,
+        out_channels: Optional[int] = 32,
+        num_attention_heads: int = 70,
+        attention_head_dim: int = 32,
+        num_layers: int = 7,
+        num_cross_attention_heads: Optional[int] = 20,
+        cross_attention_head_dim: Optional[int] = 112,
+        cross_attention_dim: Optional[int] = 2240,
+        caption_channels: int = 2304,
+        mlp_ratio: float = 2.5,
+        dropout: float = 0.0,
+        attention_bias: bool = False,
+        sample_size: int = 32,
+        patch_size: int = 1,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        interpolation_scale: Optional[int] = None,
+    ) -> None:
+        super().__init__()
+
+        out_channels = out_channels or in_channels
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # 1. Patch Embedding
+        self.patch_embed = PatchEmbed(
+            height=sample_size,
+            width=sample_size,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=inner_dim,
+            interpolation_scale=interpolation_scale,
+            pos_embed_type="sincos" if interpolation_scale is not None else None,
+        )
+
+        # 2. Additional condition embeddings
+        self.time_embed = AdaLayerNormSingle(inner_dim)
+
+        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+        self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True)
+
+        # 3. Transformer blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                SanaTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    num_cross_attention_heads=num_cross_attention_heads,
+                    cross_attention_head_dim=cross_attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_bias=attention_bias,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    mlp_ratio=mlp_ratio,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # controlnet_blocks
+        self.controlnet_blocks = nn.ModuleList([])
+
+        self.input_block = zero_module(nn.Linear(inner_dim, inner_dim))
+        for _ in range(len(self.transformer_blocks)):
+            controlnet_block = nn.Linear(inner_dim, inner_dim)
+            controlnet_block = zero_module(controlnet_block)
+            self.controlnet_blocks.append(controlnet_block)
+
+        self.gradient_checkpointing = False
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        controlnet_cond: torch.Tensor,
+        conditioning_scale: float = 1.0,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[Tuple[torch.Tensor, ...], Transformer2DModelOutput]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        batch_size, num_channels, height, width = hidden_states.shape
+        p = self.config.patch_size
+        post_patch_height, post_patch_width = height // p, width // p
+
+        hidden_states = self.patch_embed(hidden_states)
+        hidden_states = hidden_states + self.input_block(self.patch_embed(controlnet_cond.to(hidden_states.dtype)))
+
+        timestep, embedded_timestep = self.time_embed(
+            timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+        )
+
+        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        encoder_hidden_states = self.caption_norm(encoder_hidden_states)
+
+        # 2. Transformer blocks
+        block_res_samples = ()
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.transformer_blocks:
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_height,
+                    post_patch_width,
+                )
+                block_res_samples = block_res_samples + (hidden_states,)
+        else:
+            for block in self.transformer_blocks:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_height,
+                    post_patch_width,
+                )
+                block_res_samples = block_res_samples + (hidden_states,)
+
+        # 3. ControlNet blocks
+        controlnet_block_res_samples = ()
+        for block_res_sample, controlnet_block in zip(block_res_samples, self.controlnet_blocks):
+            block_res_sample = controlnet_block(block_res_sample)
+            controlnet_block_res_samples = controlnet_block_res_samples + (block_res_sample,)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        controlnet_block_res_samples = [sample * conditioning_scale for sample in controlnet_block_res_samples]
+
+        if not return_dict:
+            return (controlnet_block_res_samples,)
+
+        return SanaControlNetOutput(controlnet_block_samples=controlnet_block_res_samples)

diffusers/models/controlnets/controlnet_sd3.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved.
+# Copyright 2025 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.

diffusers/models/controlnets/controlnet_sparsectrl.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -96,7 +96,7 @@ class SparseControlNetConditioningEmbedding(nn.Module):
 class SparseControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
     """
     A SparseControlNet model as described in [SparseCtrl: Adding Sparse Controls to Text-to-Video Diffusion
-    Models](https://arxiv.org/abs/2311.16933).
+    Models](https://huggingface.co/papers/2311.16933).
 
     Args:
         in_channels (`int`, defaults to 4):

diffusers/models/controlnets/controlnet_union.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.

diffusers/models/controlnets/controlnet_xs.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -734,17 +734,17 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
             unet (`UNet2DConditionModel`):
                 The UNet model we want to control.
             controlnet (`ControlNetXSAdapter`):
-                The ConntrolNet-XS adapter with which the UNet will be fused. If none is given, a new ConntrolNet-XS
+                The ControlNet-XS adapter with which the UNet will be fused. If none is given, a new ControlNet-XS
                 adapter will be created.
             size_ratio (float, *optional*, defaults to `None`):
-                Used to contruct the controlnet if none is given. See [`ControlNetXSAdapter.from_unet`] for details.
+                Used to construct the controlnet if none is given. See [`ControlNetXSAdapter.from_unet`] for details.
             ctrl_block_out_channels (`List[int]`, *optional*, defaults to `None`):
-                Used to contruct the controlnet if none is given. See [`ControlNetXSAdapter.from_unet`] for details,
+                Used to construct the controlnet if none is given. See [`ControlNetXSAdapter.from_unet`] for details,
                 where this parameter is called `block_out_channels`.
             time_embedding_mix (`float`, *optional*, defaults to None):
-                Used to contruct the controlnet if none is given. See [`ControlNetXSAdapter.from_unet`] for details.
+                Used to construct the controlnet if none is given. See [`ControlNetXSAdapter.from_unet`] for details.
             ctrl_optional_kwargs (`Dict`, *optional*, defaults to `None`):
-                Passed to the `init` of the new controlent if no controlent was given.
+                Passed to the `init` of the new controlnet if no controlnet was given.
         """
         if controlnet is None:
             controlnet = ControlNetXSAdapter.from_unet(
@@ -942,7 +942,7 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
 
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.enable_freeu
     def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
-        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+        r"""Enables the FreeU mechanism from https://huggingface.co/papers/2309.11497.
 
         The suffixes after the scaling factors represent the stage blocks where they are being applied.
 

diffusers/models/controlnets/multicontrolnet.py

@@ -4,9 +4,9 @@ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 import torch
 from torch import nn
 
-from ...models.controlnets.controlnet import ControlNetModel, ControlNetOutput
-from ...models.modeling_utils import ModelMixin
 from ...utils import logging
+from ..controlnets.controlnet import ControlNetModel, ControlNetOutput
+from ..modeling_utils import ModelMixin
 
 
 logger = logging.get_logger(__name__)
@@ -130,9 +130,8 @@ class MultiControlNetModel(ModelMixin):
                 A path to a *directory* containing model weights saved using
                 [`~models.controlnets.multicontrolnet.MultiControlNetModel.save_pretrained`], e.g.,
                 `./my_model_directory/controlnet`.
-            torch_dtype (`str` or `torch.dtype`, *optional*):
-                Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
-                will be automatically derived from the model's weights.
+            torch_dtype (`torch.dtype`, *optional*):
+                Override the default `torch.dtype` and load the model under this dtype.
             output_loading_info(`bool`, *optional*, defaults to `False`):
                 Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
             device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):

diffusers/models/controlnets/multicontrolnet_union.py

@@ -4,10 +4,10 @@ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 import torch
 from torch import nn
 
-from ...models.controlnets.controlnet import ControlNetOutput
-from ...models.controlnets.controlnet_union import ControlNetUnionModel
-from ...models.modeling_utils import ModelMixin
 from ...utils import logging
+from ..controlnets.controlnet import ControlNetOutput
+from ..controlnets.controlnet_union import ControlNetUnionModel
+from ..modeling_utils import ModelMixin
 
 
 logger = logging.get_logger(__name__)
@@ -143,9 +143,8 @@ class MultiControlNetUnionModel(ModelMixin):
                 A path to a *directory* containing model weights saved using
                 [`~models.controlnets.multicontrolnet.MultiControlNetUnionModel.save_pretrained`], e.g.,
                 `./my_model_directory/controlnet`.
-            torch_dtype (`str` or `torch.dtype`, *optional*):
-                Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
-                will be automatically derived from the model's weights.
+            torch_dtype (`torch.dtype`, *optional*):
+                Override the default `torch.dtype` and load the model under this dtype.
             output_loading_info(`bool`, *optional*, defaults to `False`):
                 Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
             device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):

diffusers/models/downsampling.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -286,7 +286,7 @@ class KDownsample2D(nn.Module):
 
 
 class CogVideoXDownsample3D(nn.Module):
-    # Todo: Wait for paper relase.
+    # Todo: Wait for paper release.
     r"""
     A 3D Downsampling layer using in [CogVideoX]() by Tsinghua University & ZhipuAI
 

diffusers/models/embeddings.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -31,7 +31,7 @@ def get_timestep_embedding(
     downscale_freq_shift: float = 1,
     scale: float = 1,
     max_period: int = 10000,
-):
+) -> torch.Tensor:
     """
     This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.
 
@@ -97,7 +97,7 @@ def get_3d_sincos_pos_embed(
             The spatial dimension of positional embeddings. If an integer is provided, the same size is applied to both
             spatial dimensions (height and width).
         temporal_size (`int`):
-            The temporal dimension of postional embeddings (number of frames).
+            The temporal dimension of positional embeddings (number of frames).
         spatial_interpolation_scale (`float`, defaults to 1.0):
             Scale factor for spatial grid interpolation.
         temporal_interpolation_scale (`float`, defaults to 1.0):
@@ -169,7 +169,7 @@ def _get_3d_sincos_pos_embed_np(
             The spatial dimension of positional embeddings. If an integer is provided, the same size is applied to both
             spatial dimensions (height and width).
         temporal_size (`int`):
-            The temporal dimension of postional embeddings (number of frames).
+            The temporal dimension of positional embeddings (number of frames).
         spatial_interpolation_scale (`float`, defaults to 1.0):
             Scale factor for spatial grid interpolation.
         temporal_interpolation_scale (`float`, defaults to 1.0):
@@ -1149,9 +1149,7 @@ def get_1d_rotary_pos_embed(
 
     theta = theta * ntk_factor
     freqs = (
-        1.0
-        / (theta ** (torch.arange(0, dim, 2, dtype=freqs_dtype, device=pos.device)[: (dim // 2)] / dim))
-        / linear_factor
+        1.0 / (theta ** (torch.arange(0, dim, 2, dtype=freqs_dtype, device=pos.device) / dim)) / linear_factor
     )  # [D/2]
     freqs = torch.outer(pos, freqs)  # type: ignore   # [S, D/2]
     is_npu = freqs.device.type == "npu"
@@ -1201,11 +1199,11 @@ def apply_rotary_emb(
 
         if use_real_unbind_dim == -1:
             # Used for flux, cogvideox, hunyuan-dit
-            x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+            x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, H, S, D//2]
             x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
         elif use_real_unbind_dim == -2:
-            # Used for Stable Audio, OmniGen and CogView4
-            x_real, x_imag = x.reshape(*x.shape[:-1], 2, -1).unbind(-2)  # [B, S, H, D//2]
+            # Used for Stable Audio, OmniGen, CogView4 and Cosmos
+            x_real, x_imag = x.reshape(*x.shape[:-1], 2, -1).unbind(-2)  # [B, H, S, D//2]
             x_rotated = torch.cat([-x_imag, x_real], dim=-1)
         else:
             raise ValueError(f"`use_real_unbind_dim={use_real_unbind_dim}` but should be -1 or -2.")
@@ -1327,7 +1325,7 @@ class Timesteps(nn.Module):
         self.downscale_freq_shift = downscale_freq_shift
         self.scale = scale
 
-    def forward(self, timesteps):
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
         t_emb = get_timestep_embedding(
             timesteps,
             self.num_channels,
@@ -1401,7 +1399,7 @@ class ImagePositionalEmbeddings(nn.Module):
     Converts latent image classes into vector embeddings. Sums the vector embeddings with positional embeddings for the
     height and width of the latent space.
 
-    For more details, see figure 10 of the dall-e paper: https://arxiv.org/abs/2102.12092
+    For more details, see figure 10 of the dall-e paper: https://huggingface.co/papers/2102.12092
 
     For VQ-diffusion:
 

diffusers/models/embeddings_flax.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -89,7 +89,7 @@ class FlaxTimestepEmbedding(nn.Module):
 
 class FlaxTimesteps(nn.Module):
     r"""
-    Wrapper Module for sinusoidal Time step Embeddings as described in https://arxiv.org/abs/2006.11239
+    Wrapper Module for sinusoidal Time step Embeddings as described in https://huggingface.co/papers/2006.11239
 
     Args:
         dim (`int`, *optional*, defaults to `32`):

diffusers/models/lora.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -38,7 +38,7 @@ if is_transformers_available():
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-def text_encoder_attn_modules(text_encoder):
+def text_encoder_attn_modules(text_encoder: nn.Module):
     attn_modules = []
 
     if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
@@ -52,7 +52,7 @@ def text_encoder_attn_modules(text_encoder):
     return attn_modules
 
 
-def text_encoder_mlp_modules(text_encoder):
+def text_encoder_mlp_modules(text_encoder: nn.Module):
     mlp_modules = []
 
     if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):