diffusers 0.30.2__py3-none-any.whl → 0.31.0__py3-none-any.whl

diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -19,11 +19,12 @@ import torch
 from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
-from ...utils import is_torch_version, logging
+from ...loaders import PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import maybe_allow_in_graph
 from ..attention import Attention, FeedForward
 from ..attention_processor import AttentionProcessor, CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
-from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps, get_3d_sincos_pos_embed
+from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
 from ..normalization import AdaLayerNorm, CogVideoXLayerNormZero
@@ -152,7 +153,7 @@ class CogVideoXBlock(nn.Module):
         return hidden_states, encoder_hidden_states
 
 
-class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
+class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
     """
     A Transformer model for video-like data in [CogVideoX](https://github.com/THUDM/CogVideo).
 
@@ -235,37 +236,42 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
         spatial_interpolation_scale: float = 1.875,
         temporal_interpolation_scale: float = 1.0,
         use_rotary_positional_embeddings: bool = False,
+        use_learned_positional_embeddings: bool = False,
     ):
         super().__init__()
         inner_dim = num_attention_heads * attention_head_dim
 
-        post_patch_height = sample_height // patch_size
-        post_patch_width = sample_width // patch_size
-        post_time_compression_frames = (sample_frames - 1) // temporal_compression_ratio + 1
-        self.num_patches = post_patch_height * post_patch_width * post_time_compression_frames
+        if not use_rotary_positional_embeddings and use_learned_positional_embeddings:
+            raise ValueError(
+                "There are no CogVideoX checkpoints available with disable rotary embeddings and learned positional "
+                "embeddings. If you're using a custom model and/or believe this should be supported, please open an "
+                "issue at https://github.com/huggingface/diffusers/issues."
+            )
 
         # 1. Patch embedding
-        self.patch_embed = CogVideoXPatchEmbed(patch_size, in_channels, inner_dim, text_embed_dim, bias=True)
-        self.embedding_dropout = nn.Dropout(dropout)
-
-        # 2. 3D positional embeddings
-        spatial_pos_embedding = get_3d_sincos_pos_embed(
-            inner_dim,
-            (post_patch_width, post_patch_height),
-            post_time_compression_frames,
-            spatial_interpolation_scale,
-            temporal_interpolation_scale,
+        self.patch_embed = CogVideoXPatchEmbed(
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=inner_dim,
+            text_embed_dim=text_embed_dim,
+            bias=True,
+            sample_width=sample_width,
+            sample_height=sample_height,
+            sample_frames=sample_frames,
+            temporal_compression_ratio=temporal_compression_ratio,
+            max_text_seq_length=max_text_seq_length,
+            spatial_interpolation_scale=spatial_interpolation_scale,
+            temporal_interpolation_scale=temporal_interpolation_scale,
+            use_positional_embeddings=not use_rotary_positional_embeddings,
+            use_learned_positional_embeddings=use_learned_positional_embeddings,
         )
-        spatial_pos_embedding = torch.from_numpy(spatial_pos_embedding).flatten(0, 1)
-        pos_embedding = torch.zeros(1, max_text_seq_length + self.num_patches, inner_dim, requires_grad=False)
-        pos_embedding.data[:, max_text_seq_length:].copy_(spatial_pos_embedding)
-        self.register_buffer("pos_embedding", pos_embedding, persistent=False)
+        self.embedding_dropout = nn.Dropout(dropout)
 
-        # 3. Time embeddings
+        # 2. Time embeddings
         self.time_proj = Timesteps(inner_dim, flip_sin_to_cos, freq_shift)
         self.time_embedding = TimestepEmbedding(inner_dim, time_embed_dim, timestep_activation_fn)
 
-        # 4. Define spatio-temporal transformers blocks
+        # 3. Define spatio-temporal transformers blocks
         self.transformer_blocks = nn.ModuleList(
             [
                 CogVideoXBlock(
@@ -284,7 +290,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
         )
         self.norm_final = nn.LayerNorm(inner_dim, norm_eps, norm_elementwise_affine)
 
-        # 5. Output blocks
+        # 4. Output blocks
         self.norm_out = AdaLayerNorm(
             embedding_dim=time_embed_dim,
             output_dim=2 * inner_dim,
@@ -406,8 +412,24 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
         timestep: Union[int, float, torch.LongTensor],
         timestep_cond: Optional[torch.Tensor] = None,
         image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
         return_dict: bool = True,
     ):
+        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."
+                )
+
         batch_size, num_frames, channels, height, width = hidden_states.shape
 
         # 1. Time embedding
@@ -422,20 +444,13 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
 
         # 2. Patch embedding
         hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
+        hidden_states = self.embedding_dropout(hidden_states)
 
-        # 3. Position embedding
         text_seq_length = encoder_hidden_states.shape[1]
-        if not self.config.use_rotary_positional_embeddings:
-            seq_length = height * width * num_frames // (self.config.patch_size**2)
-
-            pos_embeds = self.pos_embedding[:, : text_seq_length + seq_length]
-            hidden_states = hidden_states + pos_embeds
-            hidden_states = self.embedding_dropout(hidden_states)
-
         encoder_hidden_states = hidden_states[:, :text_seq_length]
         hidden_states = hidden_states[:, text_seq_length:]
 
-        # 4. Transformer blocks
+        # 3. Transformer blocks
         for i, block in enumerate(self.transformer_blocks):
             if self.training and self.gradient_checkpointing:
 
@@ -471,15 +486,22 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
             hidden_states = self.norm_final(hidden_states)
             hidden_states = hidden_states[:, text_seq_length:]
 
-        # 5. Final block
+        # 4. Final block
         hidden_states = self.norm_out(hidden_states, temb=emb)
         hidden_states = self.proj_out(hidden_states)
 
-        # 6. Unpatchify
+        # 5. Unpatchify
+        # Note: we use `-1` instead of `channels`:
+        #   - It is okay to `channels` use for CogVideoX-2b and CogVideoX-5b (number of input channels is equal to output channels)
+        #   - However, for CogVideoX-5b-I2V also takes concatenated input image latents (number of input channels is twice the output channels)
         p = self.config.patch_size
-        output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, channels, p, p)
+        output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
         output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
 
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
         if not return_dict:
             return (output,)
         return Transformer2DModelOutput(sample=output)

diffusers/models/transformers/pixart_transformer_2d.py

@@ -19,7 +19,7 @@ from torch import nn
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import is_torch_version, logging
 from ..attention import BasicTransformerBlock
-from ..attention_processor import Attention, AttentionProcessor, FusedAttnProcessor2_0
+from ..attention_processor import Attention, AttentionProcessor, AttnProcessor, FusedAttnProcessor2_0
 from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -247,6 +247,14 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
         for name, module in self.named_children():
             fn_recursive_attn_processor(name, module, processor)
 
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+
+        Safe to just use `AttnProcessor()` as PixArt doesn't have any exotic attention processors in default model.
+        """
+        self.set_attn_processor(AttnProcessor())
+
     # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
     def fuse_qkv_projections(self):
         """

diffusers/models/transformers/transformer_cogview3plus.py

@@ -0,0 +1,386 @@
+# Copyright 2024 The CogView team, Tsinghua University & ZhipuAI 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.
+# 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 typing import Any, Dict, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...models.attention import FeedForward
+from ...models.attention_processor import (
+    Attention,
+    AttentionProcessor,
+    CogVideoXAttnProcessor2_0,
+)
+from ...models.modeling_utils import ModelMixin
+from ...models.normalization import AdaLayerNormContinuous
+from ...utils import is_torch_version, logging
+from ..embeddings import CogView3CombinedTimestepSizeEmbeddings, CogView3PlusPatchEmbed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..normalization import CogView3PlusAdaLayerNormZeroTextImage
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class CogView3PlusTransformerBlock(nn.Module):
+    r"""
+    Transformer block used in [CogView](https://github.com/THUDM/CogView3) model.
+
+    Args:
+        dim (`int`):
+            The number of channels in the input and output.
+        num_attention_heads (`int`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`):
+            The number of channels in each head.
+        time_embed_dim (`int`):
+            The number of channels in timestep embedding.
+    """
+
+    def __init__(
+        self,
+        dim: int = 2560,
+        num_attention_heads: int = 64,
+        attention_head_dim: int = 40,
+        time_embed_dim: int = 512,
+    ):
+        super().__init__()
+
+        self.norm1 = CogView3PlusAdaLayerNormZeroTextImage(embedding_dim=time_embed_dim, dim=dim)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            out_dim=dim,
+            bias=True,
+            qk_norm="layer_norm",
+            elementwise_affine=False,
+            eps=1e-6,
+            processor=CogVideoXAttnProcessor2_0(),
+        )
+
+        self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5)
+        self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5)
+
+        self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        emb: torch.Tensor,
+    ) -> torch.Tensor:
+        text_seq_length = encoder_hidden_states.size(1)
+
+        # norm & modulate
+        (
+            norm_hidden_states,
+            gate_msa,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+            norm_encoder_hidden_states,
+            c_gate_msa,
+            c_shift_mlp,
+            c_scale_mlp,
+            c_gate_mlp,
+        ) = self.norm1(hidden_states, encoder_hidden_states, emb)
+
+        # attention
+        attn_hidden_states, attn_encoder_hidden_states = self.attn1(
+            hidden_states=norm_hidden_states, encoder_hidden_states=norm_encoder_hidden_states
+        )
+
+        hidden_states = hidden_states + gate_msa.unsqueeze(1) * attn_hidden_states
+        encoder_hidden_states = encoder_hidden_states + c_gate_msa.unsqueeze(1) * attn_encoder_hidden_states
+
+        # norm & modulate
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+        norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+
+        # feed-forward
+        norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
+        ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = hidden_states + gate_mlp.unsqueeze(1) * ff_output[:, text_seq_length:]
+        encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * ff_output[:, :text_seq_length]
+
+        if hidden_states.dtype == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
+        if encoder_hidden_states.dtype == torch.float16:
+            encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
+        return hidden_states, encoder_hidden_states
+
+
+class CogView3PlusTransformer2DModel(ModelMixin, ConfigMixin):
+    r"""
+    The Transformer model introduced in [CogView3: Finer and Faster Text-to-Image Generation via Relay
+    Diffusion](https://huggingface.co/papers/2403.05121).
+
+    Args:
+        patch_size (`int`, defaults to `2`):
+            The size of the patches to use in the patch embedding layer.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        num_layers (`int`, defaults to `30`):
+            The number of layers of Transformer blocks to use.
+        attention_head_dim (`int`, defaults to `40`):
+            The number of channels in each head.
+        num_attention_heads (`int`, defaults to `64`):
+            The number of heads to use for multi-head attention.
+        out_channels (`int`, defaults to `16`):
+            The number of channels in the output.
+        text_embed_dim (`int`, defaults to `4096`):
+            Input dimension of text embeddings from the text encoder.
+        time_embed_dim (`int`, defaults to `512`):
+            Output dimension of timestep embeddings.
+        condition_dim (`int`, defaults to `256`):
+            The embedding dimension of the input SDXL-style resolution conditions (original_size, target_size,
+            crop_coords).
+        pos_embed_max_size (`int`, defaults to `128`):
+            The maximum resolution of the positional embeddings, from which slices of shape `H x W` are taken and added
+            to input patched latents, where `H` and `W` are the latent height and width respectively. A value of 128
+            means that the maximum supported height and width for image generation is `128 * vae_scale_factor *
+            patch_size => 128 * 8 * 2 => 2048`.
+        sample_size (`int`, defaults to `128`):
+            The base resolution of input latents. If height/width is not provided during generation, this value is used
+            to determine the resolution as `sample_size * vae_scale_factor => 128 * 8 => 1024`
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        num_layers: int = 30,
+        attention_head_dim: int = 40,
+        num_attention_heads: int = 64,
+        out_channels: int = 16,
+        text_embed_dim: int = 4096,
+        time_embed_dim: int = 512,
+        condition_dim: int = 256,
+        pos_embed_max_size: int = 128,
+        sample_size: int = 128,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        # CogView3 uses 3 additional SDXL-like conditions - original_size, target_size, crop_coords
+        # Each of these are sincos embeddings of shape 2 * condition_dim
+        self.pooled_projection_dim = 3 * 2 * condition_dim
+
+        self.patch_embed = CogView3PlusPatchEmbed(
+            in_channels=in_channels,
+            hidden_size=self.inner_dim,
+            patch_size=patch_size,
+            text_hidden_size=text_embed_dim,
+            pos_embed_max_size=pos_embed_max_size,
+        )
+
+        self.time_condition_embed = CogView3CombinedTimestepSizeEmbeddings(
+            embedding_dim=time_embed_dim,
+            condition_dim=condition_dim,
+            pooled_projection_dim=self.pooled_projection_dim,
+            timesteps_dim=self.inner_dim,
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                CogView3PlusTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    time_embed_dim=time_embed_dim,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        self.norm_out = AdaLayerNormContinuous(
+            embedding_dim=self.inner_dim,
+            conditioning_embedding_dim=time_embed_dim,
+            elementwise_affine=False,
+            eps=1e-6,
+        )
+        self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)
+
+        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 _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        original_size: torch.Tensor,
+        target_size: torch.Tensor,
+        crop_coords: torch.Tensor,
+        return_dict: bool = True,
+    ) -> Union[torch.Tensor, Transformer2DModelOutput]:
+        """
+        The [`CogView3PlusTransformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.Tensor`):
+                Input `hidden_states` of shape `(batch size, channel, height, width)`.
+            encoder_hidden_states (`torch.Tensor`):
+                Conditional embeddings (embeddings computed from the input conditions such as prompts) of shape
+                `(batch_size, sequence_len, text_embed_dim)`
+            timestep (`torch.LongTensor`):
+                Used to indicate denoising step.
+            original_size (`torch.Tensor`):
+                CogView3 uses SDXL-like micro-conditioning for original image size as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            target_size (`torch.Tensor`):
+                CogView3 uses SDXL-like micro-conditioning for target image size as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crop_coords (`torch.Tensor`):
+                CogView3 uses SDXL-like micro-conditioning for crop coordinates as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            `torch.Tensor` or [`~models.transformer_2d.Transformer2DModelOutput`]:
+                The denoised latents using provided inputs as conditioning.
+        """
+        height, width = hidden_states.shape[-2:]
+        text_seq_length = encoder_hidden_states.shape[1]
+
+        hidden_states = self.patch_embed(
+            hidden_states, encoder_hidden_states
+        )  # takes care of adding positional embeddings too.
+        emb = self.time_condition_embed(timestep, original_size, target_size, crop_coords, hidden_states.dtype)
+
+        encoder_hidden_states = hidden_states[:, :text_seq_length]
+        hidden_states = hidden_states[:, text_seq_length:]
+
+        for index_block, block in enumerate(self.transformer_blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states,
+                    emb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    emb=emb,
+                )
+
+        hidden_states = self.norm_out(hidden_states, emb)
+        hidden_states = self.proj_out(hidden_states)  # (batch_size, height*width, patch_size*patch_size*out_channels)
+
+        # unpatchify
+        patch_size = self.config.patch_size
+        height = height // patch_size
+        width = width // patch_size
+
+        hidden_states = hidden_states.reshape(
+            shape=(hidden_states.shape[0], height, width, self.out_channels, patch_size, patch_size)
+        )
+        hidden_states = torch.einsum("nhwcpq->nchpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(hidden_states.shape[0], self.out_channels, height * patch_size, width * patch_size)
+        )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)