unet_2d_condition.py

# Copyright 2023 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, List, Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.utils.checkpoint

from ..configuration_utils import ConfigMixin, register_to_config
from ..loaders import UNet2DConditionLoadersMixin
from ..utils import BaseOutput, logging
from .attention_processor import AttentionProcessor, AttnProcessor
from .embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
from .modeling_utils import ModelMixin
from .unet_2d_blocks import (
    CrossAttnDownBlock2D,
    CrossAttnUpBlock2D,
    DownBlock2D,
    UNetMidBlock2DCrossAttn,
    UNetMidBlock2DSimpleCrossAttn,
    UpBlock2D,
    get_down_block,
    get_up_block,
)


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


@dataclass
class UNet2DConditionOutput(BaseOutput):
    """
    Args:
        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
            Hidden states conditioned on `encoder_hidden_states` input. Output of last layer of model.
    """

    sample: torch.FloatTensor


class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
    r"""
    UNet2DConditionModel is a conditional 2D UNet model that takes in a noisy sample, conditional state, and a timestep
    and returns sample shaped output.

    This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
    implements for all the models (such as downloading or saving, etc.)

    Parameters:
        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
            Height and width of input/output sample.
        in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample.
        out_channels (`int`, *optional*, defaults to 4): The number of channels in the output.
        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
            Whether to flip the sin to cos in the time embedding.
        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
            The tuple of downsample blocks to use.
        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
            The mid block type. Choose from `UNetMidBlock2DCrossAttn` or `UNetMidBlock2DSimpleCrossAttn`, will skip the
            mid block layer if `None`.
        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D",)`):
            The tuple of upsample blocks to use.
        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
            Whether to include self-attention in the basic transformer blocks, see
            [`~models.attention.BasicTransformerBlock`].
        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
            The tuple of output channels for each block.
        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
            If `None`, it will skip the normalization and activation layers in post-processing
        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
            The dimension of the cross attention features.
        encoder_hid_dim (`int`, *optional*, defaults to None):
            If given, `encoder_hidden_states` will be projected from this dimension to `cross_attention_dim`.
        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
            for resnet blocks, see [`~models.resnet.ResnetBlock2D`]. Choose from `default` or `scale_shift`.
        class_embed_type (`str`, *optional*, defaults to None):
            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
        num_class_embeds (`int`, *optional*, defaults to None):
            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
            class conditioning with `class_embed_type` equal to `None`.
        time_embedding_type (`str`, *optional*, default to `positional`):
            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
        timestep_post_act (`str, *optional*, default to `None`):
            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
        time_cond_proj_dim (`int`, *optional*, default to `None`):
            The dimension of `cond_proj` layer in timestep embedding.
        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
        conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
        projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
            using the "projection" `class_embed_type`. Required when using the "projection" `class_embed_type`.
        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
        embeddings with the class embeddings.
    """

    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        sample_size: Optional[int] = None,
        in_channels: int = 4,
        out_channels: int = 4,
        center_input_sample: bool = False,
        flip_sin_to_cos: bool = True,
        freq_shift: int = 0,
        down_block_types: Tuple[str] = (
            "CrossAttnDownBlock2D",
            "CrossAttnDownBlock2D",
            "CrossAttnDownBlock2D",
            "DownBlock2D",
        ),
        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
        only_cross_attention: Union[bool, Tuple[bool]] = False,
        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
        layers_per_block: Union[int, Tuple[int]] = 2,
        downsample_padding: int = 1,
        mid_block_scale_factor: float = 1,
        act_fn: str = "silu",
        norm_num_groups: Optional[int] = 32,
        norm_eps: float = 1e-5,
        cross_attention_dim: Union[int, Tuple[int]] = 1280,
        encoder_hid_dim: Optional[int] = None,
        attention_head_dim: Union[int, Tuple[int]] = 8,
        dual_cross_attention: bool = False,
        use_linear_projection: bool = False,
        class_embed_type: Optional[str] = None,
        num_class_embeds: Optional[int] = None,
        upcast_attention: bool = False,
        resnet_time_scale_shift: str = "default",
        resnet_skip_time_act: bool = False,
        resnet_out_scale_factor: int = 1.0,
        time_embedding_type: str = "positional",
        timestep_post_act: Optional[str] = None,
        time_cond_proj_dim: Optional[int] = None,
        conv_in_kernel: int = 3,
        conv_out_kernel: int = 3,
        projection_class_embeddings_input_dim: Optional[int] = None,
        class_embeddings_concat: bool = False,
    ):
        super().__init__()

        self.sample_size = sample_size

        # Check inputs
        if len(down_block_types) != len(up_block_types):
            raise ValueError(
                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
            )

        if len(block_out_channels) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
            )

        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
            )

        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
            )

        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
            )

        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
            )

        # input
        conv_in_padding = (conv_in_kernel - 1) // 2
        self.conv_in = nn.Conv2d(
            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
        )

        # time
        if time_embedding_type == "fourier":
            time_embed_dim = block_out_channels[0] * 2
            if time_embed_dim % 2 != 0:
                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
            self.time_proj = GaussianFourierProjection(
                time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
            )
            timestep_input_dim = time_embed_dim
        elif time_embedding_type == "positional":
            time_embed_dim = block_out_channels[0] * 4

            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
            timestep_input_dim = block_out_channels[0]
        else:
            raise ValueError(
                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
            )

        self.time_embedding = TimestepEmbedding(
            timestep_input_dim,
            time_embed_dim,
            act_fn=act_fn,
            post_act_fn=timestep_post_act,
            cond_proj_dim=time_cond_proj_dim,
        )

        if encoder_hid_dim is not None:
            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
        else:
            self.encoder_hid_proj = None

        # class embedding
        if class_embed_type is None and num_class_embeds is not None:
            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
        elif class_embed_type == "timestep":
            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
        elif class_embed_type == "identity":
            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
        elif class_embed_type == "projection":
            if projection_class_embeddings_input_dim is None:
                raise ValueError(
                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
                )
            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
            # 2. it projects from an arbitrary input dimension.
            #
            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
        elif class_embed_type == "simple_projection":
            if projection_class_embeddings_input_dim is None:
                raise ValueError(
                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
                )
            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
        else:
            self.class_embedding = None

        self.down_blocks = nn.ModuleList([])
        self.up_blocks = nn.ModuleList([])

        if isinstance(only_cross_attention, bool):
            only_cross_attention = [only_cross_attention] * len(down_block_types)

        if isinstance(attention_head_dim, int):
            attention_head_dim = (attention_head_dim,) * len(down_block_types)

        if isinstance(cross_attention_dim, int):
            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)

        if isinstance(layers_per_block, int):
            layers_per_block = [layers_per_block] * len(down_block_types)

        if class_embeddings_concat:
            # The time embeddings are concatenated with the class embeddings. The dimension of the
            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
            # regular time embeddings
            blocks_time_embed_dim = time_embed_dim * 2
        else:
            blocks_time_embed_dim = time_embed_dim

        # down
        output_channel = block_out_channels[0]
        for i, down_block_type in enumerate(down_block_types):
            input_channel = output_channel
            output_channel = block_out_channels[i]
            is_final_block = i == len(block_out_channels) - 1

            down_block = get_down_block(
                down_block_type,
                num_layers=layers_per_block[i],
                in_channels=input_channel,
                out_channels=output_channel,
                temb_channels=blocks_time_embed_dim,
                add_downsample=not is_final_block,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                resnet_groups=norm_num_groups,
                cross_attention_dim=cross_attention_dim[i],
                attn_num_head_channels=attention_head_dim[i],
                downsample_padding=downsample_padding,
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                only_cross_attention=only_cross_attention[i],
                upcast_attention=upcast_attention,
                resnet_time_scale_shift=resnet_time_scale_shift,
                resnet_skip_time_act=resnet_skip_time_act,
                resnet_out_scale_factor=resnet_out_scale_factor,
            )
            self.down_blocks.append(down_block)

        # mid
        if mid_block_type == "UNetMidBlock2DCrossAttn":
            self.mid_block = UNetMidBlock2DCrossAttn(
                in_channels=block_out_channels[-1],
                temb_channels=blocks_time_embed_dim,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
                resnet_time_scale_shift=resnet_time_scale_shift,
                cross_attention_dim=cross_attention_dim[-1],
                attn_num_head_channels=attention_head_dim[-1],
                resnet_groups=norm_num_groups,
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                upcast_attention=upcast_attention,
            )
        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
                in_channels=block_out_channels[-1],
                temb_channels=blocks_time_embed_dim,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
                cross_attention_dim=cross_attention_dim[-1],
                attn_num_head_channels=attention_head_dim[-1],
                resnet_groups=norm_num_groups,
                resnet_time_scale_shift=resnet_time_scale_shift,
                skip_time_act=resnet_skip_time_act,
            )
        elif mid_block_type is None:
            self.mid_block = None
        else:
            raise ValueError(f"unknown mid_block_type : {mid_block_type}")

        # count how many layers upsample the images
        self.num_upsamplers = 0

        # up
        reversed_block_out_channels = list(reversed(block_out_channels))
        reversed_attention_head_dim = list(reversed(attention_head_dim))
        reversed_layers_per_block = list(reversed(layers_per_block))
        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
        only_cross_attention = list(reversed(only_cross_attention))

        output_channel = reversed_block_out_channels[0]
        for i, up_block_type in enumerate(up_block_types):
            is_final_block = i == len(block_out_channels) - 1

            prev_output_channel = output_channel
            output_channel = reversed_block_out_channels[i]
            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]

            # add upsample block for all BUT final layer
            if not is_final_block:
                add_upsample = True
                self.num_upsamplers += 1
            else:
                add_upsample = False

            up_block = get_up_block(
                up_block_type,
                num_layers=reversed_layers_per_block[i] + 1,
                in_channels=input_channel,
                out_channels=output_channel,
                prev_output_channel=prev_output_channel,
                temb_channels=blocks_time_embed_dim,
                add_upsample=add_upsample,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                resnet_groups=norm_num_groups,
                cross_attention_dim=reversed_cross_attention_dim[i],
                attn_num_head_channels=reversed_attention_head_dim[i],
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                only_cross_attention=only_cross_attention[i],
                upcast_attention=upcast_attention,
                resnet_time_scale_shift=resnet_time_scale_shift,
                resnet_skip_time_act=resnet_skip_time_act,
                resnet_out_scale_factor=resnet_out_scale_factor,
            )
            self.up_blocks.append(up_block)
            prev_output_channel = output_channel

        # out
        if norm_num_groups is not None:
            self.conv_norm_out = nn.GroupNorm(
                num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
            )
            self.conv_act = nn.SiLU()
        else:
            self.conv_norm_out = None
            self.conv_act = None

        conv_out_padding = (conv_out_kernel - 1) // 2
        self.conv_out = nn.Conv2d(
            block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
        )

    @property
    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, "set_processor"):
                processors[f"{name}.processor"] = module.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

    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
        r"""
        Parameters:
            `processor (`dict` of `AttentionProcessor` or `AttentionProcessor`):
                The instantiated processor class or a dictionary of processor classes that will be set as the processor
                of **all** `Attention` layers.
            In case `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_default_attn_processor(self):
        """
        Disables custom attention processors and sets the default attention implementation.
        """
        self.set_attn_processor(AttnProcessor())

    def set_attention_slice(self, slice_size):
        r"""
        Enable sliced attention computation.

        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
        in several steps. This is useful to save some memory in exchange for a small speed decrease.

        Args:
            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
                `"max"`, maximum amount of memory will be saved by running only one slice at a time. If a number is
                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
                must be a multiple of `slice_size`.
        """
        sliceable_head_dims = []

        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
            if hasattr(module, "set_attention_slice"):
                sliceable_head_dims.append(module.sliceable_head_dim)

            for child in module.children():
                fn_recursive_retrieve_sliceable_dims(child)

        # retrieve number of attention layers
        for module in self.children():
            fn_recursive_retrieve_sliceable_dims(module)

        num_sliceable_layers = len(sliceable_head_dims)

        if slice_size == "auto":
            # half the attention head size is usually a good trade-off between
            # speed and memory
            slice_size = [dim // 2 for dim in sliceable_head_dims]
        elif slice_size == "max":
            # make smallest slice possible
            slice_size = num_sliceable_layers * [1]

        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size

        if len(slice_size) != len(sliceable_head_dims):
            raise ValueError(
                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
            )

        for i in range(len(slice_size)):
            size = slice_size[i]
            dim = sliceable_head_dims[i]
            if size is not None and size > dim:
                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")

        # Recursively walk through all the children.
        # Any children which exposes the set_attention_slice method
        # gets the message
        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
            if hasattr(module, "set_attention_slice"):
                module.set_attention_slice(slice_size.pop())

            for child in module.children():
                fn_recursive_set_attention_slice(child, slice_size)

        reversed_slice_size = list(reversed(slice_size))
        for module in self.children():
            fn_recursive_set_attention_slice(module, reversed_slice_size)

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D, CrossAttnUpBlock2D, UpBlock2D)):
            module.gradient_checkpointing = value

    def forward(
        self,
        sample: torch.FloatTensor,
        timestep: Union[torch.Tensor, float, int],
        encoder_hidden_states: torch.Tensor,
        class_labels: Optional[torch.Tensor] = None,
        timestep_cond: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
        mid_block_additional_residual: Optional[torch.Tensor] = None,
        return_dict: bool = True,
    ) -> Union[UNet2DConditionOutput, Tuple]:
        r"""
        Args:
            sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
            timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
            encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).

        Returns:
            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
            [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
            returning a tuple, the first element is the sample tensor.
        """
        # By default samples have to be AT least a multiple of the overall upsampling factor.
        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
        # However, the upsampling interpolation output size can be forced to fit any upsampling size
        # on the fly if necessary.
        default_overall_up_factor = 2**self.num_upsamplers

        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
        forward_upsample_size = False
        upsample_size = None

        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
            logger.info("Forward upsample size to force interpolation output size.")
            forward_upsample_size = True

        # prepare attention_mask
        if attention_mask is not None:
            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
            attention_mask = attention_mask.unsqueeze(1)

        # 0. center input if necessary
        if self.config.center_input_sample:
            sample = 2 * sample - 1.0

        # 1. time
        timesteps = timestep
        if not torch.is_tensor(timesteps):
            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
            # This would be a good case for the `match` statement (Python 3.10+)
            is_mps = sample.device.type == "mps"
            if isinstance(timestep, float):
                dtype = torch.float32 if is_mps else torch.float64
            else:
                dtype = torch.int32 if is_mps else torch.int64
            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
        elif len(timesteps.shape) == 0:
            timesteps = timesteps[None].to(sample.device)

        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
        timesteps = timesteps.expand(sample.shape[0])

        t_emb = self.time_proj(timesteps)

        # timesteps does not contain any weights and will always return f32 tensors
        # but time_embedding might actually be running in fp16. so we need to cast here.
        # there might be better ways to encapsulate this.
        t_emb = t_emb.to(dtype=self.dtype)

        emb = self.time_embedding(t_emb, timestep_cond)

        if self.class_embedding is not None:
            if class_labels is None:
                raise ValueError("class_labels should be provided when num_class_embeds > 0")

            if self.config.class_embed_type == "timestep":
                class_labels = self.time_proj(class_labels)

            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)

            if self.config.class_embeddings_concat:
                emb = torch.cat([emb, class_emb], dim=-1)
            else:
                emb = emb + class_emb

        if self.encoder_hid_proj is not None:
            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)

        # 2. pre-process
        sample = self.conv_in(sample)

        # 3. down
        down_block_res_samples = (sample,)
        for downsample_block in self.down_blocks:
            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
                sample, res_samples = downsample_block(
                    hidden_states=sample,
                    temb=emb,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                )
            else:
                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)

            down_block_res_samples += res_samples

        if down_block_additional_residuals is not None:
            new_down_block_res_samples = ()

            for down_block_res_sample, down_block_additional_residual in zip(
                down_block_res_samples, down_block_additional_residuals
            ):
                down_block_res_sample = down_block_res_sample + down_block_additional_residual
                new_down_block_res_samples += (down_block_res_sample,)

            down_block_res_samples = new_down_block_res_samples

        # 4. mid
        if self.mid_block is not None:
            sample = self.mid_block(
                sample,
                emb,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=attention_mask,
                cross_attention_kwargs=cross_attention_kwargs,
            )

        if mid_block_additional_residual is not None:
            sample = sample + mid_block_additional_residual

        # 5. up
        for i, upsample_block in enumerate(self.up_blocks):
            is_final_block = i == len(self.up_blocks) - 1

            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]

            # if we have not reached the final block and need to forward the
            # upsample size, we do it here
            if not is_final_block and forward_upsample_size:
                upsample_size = down_block_res_samples[-1].shape[2:]

            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
                sample = upsample_block(
                    hidden_states=sample,
                    temb=emb,
                    res_hidden_states_tuple=res_samples,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    upsample_size=upsample_size,
                    attention_mask=attention_mask,
                )
            else:
                sample = upsample_block(
                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
                )

        # 6. post-process
        if self.conv_norm_out:
            sample = self.conv_norm_out(sample)
            sample = self.conv_act(sample)
        sample = self.conv_out(sample)

        if not return_dict:
            return (sample,)

        return UNet2DConditionOutput(sample=sample)