modeling_text_unet.py

from typing import Any, Dict, List, Optional, Tuple, Union

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from ...configuration_utils import ConfigMixin, register_to_config
from ...models import ModelMixin
from ...models.activations import get_activation
from ...models.attention import Attention
from ...models.attention_processor import (
    AttentionProcessor,
    AttnAddedKVProcessor,
    AttnAddedKVProcessor2_0,
    AttnProcessor,
)
from ...models.dual_transformer_2d import DualTransformer2DModel
from ...models.embeddings import (
    GaussianFourierProjection,
    TextImageProjection,
    TextImageTimeEmbedding,
    TextTimeEmbedding,
    TimestepEmbedding,
    Timesteps,
)
from ...models.transformer_2d import Transformer2DModel
from ...models.unet_2d_condition import UNet2DConditionOutput
from ...utils import is_torch_version, logging


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


def get_down_block(
    down_block_type,
    num_layers,
    in_channels,
    out_channels,
    temb_channels,
    add_downsample,
    resnet_eps,
    resnet_act_fn,
    num_attention_heads,
    resnet_groups=None,
    cross_attention_dim=None,
    downsample_padding=None,
    dual_cross_attention=False,
    use_linear_projection=False,
    only_cross_attention=False,
    upcast_attention=False,
    resnet_time_scale_shift="default",
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
    cross_attention_norm=None,
):
    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
    if down_block_type == "DownBlockFlat":
        return DownBlockFlat(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "CrossAttnDownBlockFlat":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockFlat")
        return CrossAttnDownBlockFlat(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            cross_attention_dim=cross_attention_dim,
            num_attention_heads=num_attention_heads,
            dual_cross_attention=dual_cross_attention,
            use_linear_projection=use_linear_projection,
            only_cross_attention=only_cross_attention,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    raise ValueError(f"{down_block_type} is not supported.")


def get_up_block(
    up_block_type,
    num_layers,
    in_channels,
    out_channels,
    prev_output_channel,
    temb_channels,
    add_upsample,
    resnet_eps,
    resnet_act_fn,
    num_attention_heads,
    resnet_groups=None,
    cross_attention_dim=None,
    dual_cross_attention=False,
    use_linear_projection=False,
    only_cross_attention=False,
    upcast_attention=False,
    resnet_time_scale_shift="default",
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
    cross_attention_norm=None,
):
    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
    if up_block_type == "UpBlockFlat":
        return UpBlockFlat(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "CrossAttnUpBlockFlat":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockFlat")
        return CrossAttnUpBlockFlat(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
            num_attention_heads=num_attention_heads,
            dual_cross_attention=dual_cross_attention,
            use_linear_projection=use_linear_projection,
            only_cross_attention=only_cross_attention,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    raise ValueError(f"{up_block_type} is not supported.")


# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel with UNet2DConditionModel->UNetFlatConditionModel, nn.Conv2d->LinearMultiDim, Block2D->BlockFlat
class UNetFlatConditionModel(ModelMixin, ConfigMixin):
    r"""
    UNetFlatConditionModel 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 `("CrossAttnDownBlockFlat", "CrossAttnDownBlockFlat", "CrossAttnDownBlockFlat", "DownBlockFlat")`):
            The tuple of downsample blocks to use.
        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlockFlatCrossAttn"`):
            The mid block type. Choose from `UNetMidBlockFlatCrossAttn` or `UNetMidBlockFlatSimpleCrossAttn`, will skip
            the mid block layer if `None`.
        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlockFlat", "CrossAttnUpBlockFlat", "CrossAttnUpBlockFlat", "CrossAttnUpBlockFlat",)`):
            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 `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
            dimension to `cross_attention_dim`.
        encoder_hid_dim_type (`str`, *optional*, defaults to None):
            If given, the `encoder_hidden_states` and potentially other embeddings will be down-projected to text
            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
        num_attention_heads (`int`, *optional*):
            The number of attention heads. If not defined, defaults to `attention_head_dim`
        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
            for resnet blocks, see [`~models.resnet.ResnetBlockFlat`]. 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"`.
        addition_embed_type (`str`, *optional*, defaults to None):
            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
            "text". "text" will use the `TextTimeEmbedding` layer.
        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`.
        time_embedding_dim (`int`, *optional*, default to `None`):
            An optional override for the dimension of the projected time embedding.
        time_embedding_act_fn (`str`, *optional*, default to `None`):
            Optional activation function to use on the time embeddings only one time before they as passed to the rest
            of the unet. Choose from `silu`, `mish`, `gelu`, and `swish`.
        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.
        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
            Whether to use cross attention with the mid block when using the `UNetMidBlockFlatSimpleCrossAttn`. If
            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is None, the
            `only_cross_attention` value will be used as the value for `mid_block_only_cross_attention`. Else, it will
            default to `False`.
    """

    _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] = (
            "CrossAttnDownBlockFlat",
            "CrossAttnDownBlockFlat",
            "CrossAttnDownBlockFlat",
            "DownBlockFlat",
        ),
        mid_block_type: Optional[str] = "UNetMidBlockFlatCrossAttn",
        up_block_types: Tuple[str] = (
            "UpBlockFlat",
            "CrossAttnUpBlockFlat",
            "CrossAttnUpBlockFlat",
            "CrossAttnUpBlockFlat",
        ),
        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,
        encoder_hid_dim_type: Optional[str] = None,
        attention_head_dim: Union[int, Tuple[int]] = 8,
        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
        dual_cross_attention: bool = False,
        use_linear_projection: bool = False,
        class_embed_type: Optional[str] = None,
        addition_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",
        time_embedding_dim: Optional[int] = None,
        time_embedding_act_fn: Optional[str] = None,
        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,
        mid_block_only_cross_attention: Optional[bool] = None,
        cross_attention_norm: Optional[str] = None,
        addition_embed_type_num_heads=64,
    ):
        super().__init__()

        self.sample_size = sample_size

        # If `num_attention_heads` is not defined (which is the case for most models)
        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
        # The reason for this behavior is to correct for incorrectly named variables that were introduced
        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
        # which is why we correct for the naming here.
        num_attention_heads = num_attention_heads or attention_head_dim

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

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

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

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

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

        # time
        if time_embedding_type == "fourier":
            time_embed_dim = time_embedding_dim or 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 = time_embedding_dim or 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_type is None and encoder_hid_dim is not None:
            encoder_hid_dim_type = "text_proj"
            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")

        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
            raise ValueError(
                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
            )

        if encoder_hid_dim_type == "text_proj":
            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
        elif encoder_hid_dim_type == "text_image_proj":
            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
            self.encoder_hid_proj = TextImageProjection(
                text_embed_dim=encoder_hid_dim,
                image_embed_dim=cross_attention_dim,
                cross_attention_dim=cross_attention_dim,
            )

        elif encoder_hid_dim_type is not None:
            raise ValueError(
                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
            )
        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, act_fn=act_fn)
        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

        if addition_embed_type == "text":
            if encoder_hid_dim is not None:
                text_time_embedding_from_dim = encoder_hid_dim
            else:
                text_time_embedding_from_dim = cross_attention_dim

            self.add_embedding = TextTimeEmbedding(
                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
            )
        elif addition_embed_type == "text_image":
            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
            self.add_embedding = TextImageTimeEmbedding(
                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
            )
        elif addition_embed_type is not None:
            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")

        if time_embedding_act_fn is None:
            self.time_embed_act = None
        else:
            self.time_embed_act = get_activation(time_embedding_act_fn)

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

        if isinstance(only_cross_attention, bool):
            if mid_block_only_cross_attention is None:
                mid_block_only_cross_attention = only_cross_attention

            only_cross_attention = [only_cross_attention] * len(down_block_types)

        if mid_block_only_cross_attention is None:
            mid_block_only_cross_attention = False

        if isinstance(num_attention_heads, int):
            num_attention_heads = (num_attention_heads,) * 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],
                num_attention_heads=num_attention_heads[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,
                cross_attention_norm=cross_attention_norm,
                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
            )
            self.down_blocks.append(down_block)

        # mid
        if mid_block_type == "UNetMidBlockFlatCrossAttn":
            self.mid_block = UNetMidBlockFlatCrossAttn(
                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],
                num_attention_heads=num_attention_heads[-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 == "UNetMidBlockFlatSimpleCrossAttn":
            self.mid_block = UNetMidBlockFlatSimpleCrossAttn(
                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],
                attention_head_dim=attention_head_dim[-1],
                resnet_groups=norm_num_groups,
                resnet_time_scale_shift=resnet_time_scale_shift,
                skip_time_act=resnet_skip_time_act,
                only_cross_attention=mid_block_only_cross_attention,
                cross_attention_norm=cross_attention_norm,
            )
        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_num_attention_heads = list(reversed(num_attention_heads))
        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],
                num_attention_heads=reversed_num_attention_heads[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,
                cross_attention_norm=cross_attention_norm,
                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
            )
            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 = get_activation(act_fn)

        else:
            self.conv_norm_out = None
            self.conv_act = None

        conv_out_padding = (conv_out_kernel - 1) // 2
        self.conv_out = LinearMultiDim(
            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 `num_attention_heads // slice_size`. In this case,
                `num_attention_heads` 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, (CrossAttnDownBlockFlat, DownBlockFlat, CrossAttnUpBlockFlat, UpBlockFlat)):
            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,
        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
        mid_block_additional_residual: Optional[torch.Tensor] = None,
        encoder_attention_mask: 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
            encoder_attention_mask (`torch.Tensor`):
                (batch, sequence_length) cross-attention mask, applied to encoder_hidden_states. True = keep, False =
                discard. Mask will be converted into a bias, which adds large negative values to attention scores
                corresponding to "discard" tokens.
            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).
            added_cond_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified includes additonal conditions that can be used for additonal time
                embeddings or encoder hidden states projections. See the configurations `encoder_hid_dim_type` and
                `addition_embed_type` for more information.

        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

        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
        # 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:
            # 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(sample.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:
            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
            encoder_attention_mask = encoder_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=sample.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)

                # `Timesteps` does not contain any weights and will always return f32 tensors
                # there might be better ways to encapsulate this.
                class_labels = class_labels.to(dtype=sample.dtype)

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

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

        if self.config.addition_embed_type == "text":
            aug_emb = self.add_embedding(encoder_hidden_states)
            emb = emb + aug_emb
        elif self.config.addition_embed_type == "text_image":
            # Kadinsky 2.1 - style
            if "image_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires"
                    " the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
                )

            image_embs = added_cond_kwargs.get("image_embeds")
            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)

            aug_emb = self.add_embedding(text_embs, image_embs)
            emb = emb + aug_emb

        if self.time_embed_act is not None:
            emb = self.time_embed_act(emb)

        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
            # Kadinsky 2.1 - style
            if "image_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which"
                    " requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
                )

            image_embeds = added_cond_kwargs.get("image_embeds")
            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)

        # 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,
                    encoder_attention_mask=encoder_attention_mask,
                )
            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 = 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,
                encoder_attention_mask=encoder_attention_mask,
            )

        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,
                    encoder_attention_mask=encoder_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)


class LinearMultiDim(nn.Linear):
    def __init__(self, in_features, out_features=None, second_dim=4, *args, **kwargs):
        in_features = [in_features, second_dim, 1] if isinstance(in_features, int) else list(in_features)
        if out_features is None:
            out_features = in_features
        out_features = [out_features, second_dim, 1] if isinstance(out_features, int) else list(out_features)
        self.in_features_multidim = in_features
        self.out_features_multidim = out_features
        super().__init__(np.array(in_features).prod(), np.array(out_features).prod())

    def forward(self, input_tensor, *args, **kwargs):
        shape = input_tensor.shape
        n_dim = len(self.in_features_multidim)
        input_tensor = input_tensor.reshape(*shape[0:-n_dim], self.in_features)
        output_tensor = super().forward(input_tensor)
        output_tensor = output_tensor.view(*shape[0:-n_dim], *self.out_features_multidim)
        return output_tensor


class ResnetBlockFlat(nn.Module):
    def __init__(
        self,
        *,
        in_channels,
        out_channels=None,
        dropout=0.0,
        temb_channels=512,
        groups=32,
        groups_out=None,
        pre_norm=True,
        eps=1e-6,
        time_embedding_norm="default",
        use_in_shortcut=None,
        second_dim=4,
        **kwargs,
    ):
        super().__init__()
        self.pre_norm = pre_norm
        self.pre_norm = True

        in_channels = [in_channels, second_dim, 1] if isinstance(in_channels, int) else list(in_channels)
        self.in_channels_prod = np.array(in_channels).prod()
        self.channels_multidim = in_channels

        if out_channels is not None:
            out_channels = [out_channels, second_dim, 1] if isinstance(out_channels, int) else list(out_channels)
            out_channels_prod = np.array(out_channels).prod()
            self.out_channels_multidim = out_channels
        else:
            out_channels_prod = self.in_channels_prod
            self.out_channels_multidim = self.channels_multidim
        self.time_embedding_norm = time_embedding_norm

        if groups_out is None:
            groups_out = groups

        self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=self.in_channels_prod, eps=eps, affine=True)
        self.conv1 = torch.nn.Conv2d(self.in_channels_prod, out_channels_prod, kernel_size=1, padding=0)

        if temb_channels is not None:
            self.time_emb_proj = torch.nn.Linear(temb_channels, out_channels_prod)
        else:
            self.time_emb_proj = None

        self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels_prod, eps=eps, affine=True)
        self.dropout = torch.nn.Dropout(dropout)
        self.conv2 = torch.nn.Conv2d(out_channels_prod, out_channels_prod, kernel_size=1, padding=0)

        self.nonlinearity = nn.SiLU()

        self.use_in_shortcut = (
            self.in_channels_prod != out_channels_prod if use_in_shortcut is None else use_in_shortcut
        )

        self.conv_shortcut = None
        if self.use_in_shortcut:
            self.conv_shortcut = torch.nn.Conv2d(
                self.in_channels_prod, out_channels_prod, kernel_size=1, stride=1, padding=0
            )

    def forward(self, input_tensor, temb):
        shape = input_tensor.shape
        n_dim = len(self.channels_multidim)
        input_tensor = input_tensor.reshape(*shape[0:-n_dim], self.in_channels_prod, 1, 1)
        input_tensor = input_tensor.view(-1, self.in_channels_prod, 1, 1)

        hidden_states = input_tensor

        hidden_states = self.norm1(hidden_states)
        hidden_states = self.nonlinearity(hidden_states)
        hidden_states = self.conv1(hidden_states)

        if temb is not None:
            temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None]
            hidden_states = hidden_states + temb

        hidden_states = self.norm2(hidden_states)
        hidden_states = self.nonlinearity(hidden_states)

        hidden_states = self.dropout(hidden_states)
        hidden_states = self.conv2(hidden_states)

        if self.conv_shortcut is not None:
            input_tensor = self.conv_shortcut(input_tensor)

        output_tensor = input_tensor + hidden_states

        output_tensor = output_tensor.view(*shape[0:-n_dim], -1)
        output_tensor = output_tensor.view(*shape[0:-n_dim], *self.out_channels_multidim)

        return output_tensor


# Copied from diffusers.models.unet_2d_blocks.DownBlock2D with DownBlock2D->DownBlockFlat, ResnetBlock2D->ResnetBlockFlat, Downsample2D->LinearMultiDim
class DownBlockFlat(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
                ResnetBlockFlat(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    LinearMultiDim(
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    def forward(self, hidden_states, temb=None):
        output_states = ()

        for resnet in self.resnets:
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
            else:
                hidden_states = resnet(hidden_states, temb)

            output_states = output_states + (hidden_states,)

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
                hidden_states = downsampler(hidden_states)

            output_states = output_states + (hidden_states,)

        return hidden_states, output_states


# Copied from diffusers.models.unet_2d_blocks.CrossAttnDownBlock2D with CrossAttnDownBlock2D->CrossAttnDownBlockFlat, ResnetBlock2D->ResnetBlockFlat, Downsample2D->LinearMultiDim
class CrossAttnDownBlockFlat(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        num_attention_heads=1,
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        downsample_padding=1,
        add_downsample=True,
        dual_cross_attention=False,
        use_linear_projection=False,
        only_cross_attention=False,
        upcast_attention=False,
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
                ResnetBlockFlat(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                        use_linear_projection=use_linear_projection,
                        only_cross_attention=only_cross_attention,
                        upcast_attention=upcast_attention,
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
                )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    LinearMultiDim(
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
    ):
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    None,  # timestep
                    None,  # class_labels
                    cross_attention_kwargs,
                    attention_mask,
                    encoder_attention_mask,
                    **ckpt_kwargs,
                )[0]
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
                )[0]

            output_states = output_states + (hidden_states,)

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
                hidden_states = downsampler(hidden_states)

            output_states = output_states + (hidden_states,)

        return hidden_states, output_states


# Copied from diffusers.models.unet_2d_blocks.UpBlock2D with UpBlock2D->UpBlockFlat, ResnetBlock2D->ResnetBlockFlat, Upsample2D->LinearMultiDim
class UpBlockFlat(nn.Module):
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            resnets.append(
                ResnetBlockFlat(
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([LinearMultiDim(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
            else:
                hidden_states = resnet(hidden_states, temb)

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
                hidden_states = upsampler(hidden_states, upsample_size)

        return hidden_states


# Copied from diffusers.models.unet_2d_blocks.CrossAttnUpBlock2D with CrossAttnUpBlock2D->CrossAttnUpBlockFlat, ResnetBlock2D->ResnetBlockFlat, Upsample2D->LinearMultiDim
class CrossAttnUpBlockFlat(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        num_attention_heads=1,
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
        dual_cross_attention=False,
        use_linear_projection=False,
        only_cross_attention=False,
        upcast_attention=False,
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads

        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            resnets.append(
                ResnetBlockFlat(
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                        use_linear_projection=use_linear_projection,
                        only_cross_attention=only_cross_attention,
                        upcast_attention=upcast_attention,
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
                )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([LinearMultiDim(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        upsample_size: Optional[int] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
    ):
        for resnet, attn in zip(self.resnets, self.attentions):
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    None,  # timestep
                    None,  # class_labels
                    cross_attention_kwargs,
                    attention_mask,
                    encoder_attention_mask,
                    **ckpt_kwargs,
                )[0]
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
                )[0]

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
                hidden_states = upsampler(hidden_states, upsample_size)

        return hidden_states


# Copied from diffusers.models.unet_2d_blocks.UNetMidBlock2DCrossAttn with UNetMidBlock2DCrossAttn->UNetMidBlockFlatCrossAttn, ResnetBlock2D->ResnetBlockFlat
class UNetMidBlockFlatCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        num_attention_heads=1,
        output_scale_factor=1.0,
        cross_attention_dim=1280,
        dual_cross_attention=False,
        use_linear_projection=False,
        upcast_attention=False,
    ):
        super().__init__()

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

        # there is always at least one resnet
        resnets = [
            ResnetBlockFlat(
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=resnet_groups,
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
            )
        ]
        attentions = []

        for _ in range(num_layers):
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
                        num_attention_heads,
                        in_channels // num_attention_heads,
                        in_channels=in_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                        use_linear_projection=use_linear_projection,
                        upcast_attention=upcast_attention,
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
                        num_attention_heads,
                        in_channels // num_attention_heads,
                        in_channels=in_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
                )
            resnets.append(
                ResnetBlockFlat(
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
    ) -> torch.FloatTensor:
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            hidden_states = attn(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                cross_attention_kwargs=cross_attention_kwargs,
                attention_mask=attention_mask,
                encoder_attention_mask=encoder_attention_mask,
                return_dict=False,
            )[0]
            hidden_states = resnet(hidden_states, temb)

        return hidden_states


# Copied from diffusers.models.unet_2d_blocks.UNetMidBlock2DSimpleCrossAttn with UNetMidBlock2DSimpleCrossAttn->UNetMidBlockFlatSimpleCrossAttn, ResnetBlock2D->ResnetBlockFlat
class UNetMidBlockFlatSimpleCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        attention_head_dim=1,
        output_scale_factor=1.0,
        cross_attention_dim=1280,
        skip_time_act=False,
        only_cross_attention=False,
        cross_attention_norm=None,
    ):
        super().__init__()

        self.has_cross_attention = True

        self.attention_head_dim = attention_head_dim
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

        self.num_heads = in_channels // self.attention_head_dim

        # there is always at least one resnet
        resnets = [
            ResnetBlockFlat(
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=resnet_groups,
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
                skip_time_act=skip_time_act,
            )
        ]
        attentions = []

        for _ in range(num_layers):
            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

            attentions.append(
                Attention(
                    query_dim=in_channels,
                    cross_attention_dim=in_channels,
                    heads=self.num_heads,
                    dim_head=self.attention_head_dim,
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
                    only_cross_attention=only_cross_attention,
                    cross_attention_norm=cross_attention_norm,
                    processor=processor,
                )
            )
            resnets.append(
                ResnetBlockFlat(
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                    skip_time_act=skip_time_act,
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}

        if attention_mask is None:
            # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
            mask = None if encoder_hidden_states is None else encoder_attention_mask
        else:
            # when attention_mask is defined: we don't even check for encoder_attention_mask.
            # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
            # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
            #       then we can simplify this whole if/else block to:
            #         mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
            mask = attention_mask

        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            # attn
            hidden_states = attn(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=mask,
                **cross_attention_kwargs,
            )

            # resnet
            hidden_states = resnet(hidden_states, temb)

        return hidden_states