src/diffusers/pipelines/unidiffuser/modeling_uvit.py

import math
from typing import Optional, Union

import torch
from torch import nn

from ...configuration_utils import ConfigMixin, register_to_config
from ...models import ModelMixin
from ...models.attention import FeedForward
from ...models.attention_processor import Attention
from ...models.embeddings import TimestepEmbedding, Timesteps, get_2d_sincos_pos_embed
from ...models.modeling_outputs import Transformer2DModelOutput
from ...models.normalization import AdaLayerNorm
from ...utils import logging


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


def _no_grad_trunc_normal_(tensor, mean, std, a, b):
    # Cut & paste from PyTorch official master until it's in a few official releases - RW
    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
    def norm_cdf(x):
        # Computes standard normal cumulative distribution function
        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0

    if (mean < a - 2 * std) or (mean > b + 2 * std):
        logger.warning(
            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
            "The distribution of values may be incorrect."
        )

    with torch.no_grad():
        # Values are generated by using a truncated uniform distribution and
        # then using the inverse CDF for the normal distribution.
        # Get upper and lower cdf values
        l = norm_cdf((a - mean) / std)
        u = norm_cdf((b - mean) / std)

        # Uniformly fill tensor with values from [l, u], then translate to
        # [2l-1, 2u-1].
        tensor.uniform_(2 * l - 1, 2 * u - 1)

        # Use inverse cdf transform for normal distribution to get truncated
        # standard normal
        tensor.erfinv_()

        # Transform to proper mean, std
        tensor.mul_(std * math.sqrt(2.0))
        tensor.add_(mean)

        # Clamp to ensure it's in the proper range
        tensor.clamp_(min=a, max=b)
        return tensor


def trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):
    # type: (torch.Tensor, float, float, float, float) -> torch.Tensor
    r"""Fills the input Tensor with values drawn from a truncated
    normal distribution. The values are effectively drawn from the normal distribution :math:`\mathcal{N}(\text{mean},
    \text{std}^2)` with values outside :math:`[a, b]` redrawn until they are within the bounds. The method used for
    generating the random values works best when :math:`a \leq \text{mean} \leq b`.

    Args:
        tensor: an n-dimensional `torch.Tensor`
        mean: the mean of the normal distribution
        std: the standard deviation of the normal distribution
        a: the minimum cutoff value
        b: the maximum cutoff value
    Examples:
        >>> w = torch.empty(3, 5) >>> nn.init.trunc_normal_(w)
    """
    return _no_grad_trunc_normal_(tensor, mean, std, a, b)


class PatchEmbed(nn.Module):
    """2D Image to Patch Embedding"""

    def __init__(
        self,
        height=224,
        width=224,
        patch_size=16,
        in_channels=3,
        embed_dim=768,
        layer_norm=False,
        flatten=True,
        bias=True,
        use_pos_embed=True,
    ):
        super().__init__()

        num_patches = (height // patch_size) * (width // patch_size)
        self.flatten = flatten
        self.layer_norm = layer_norm

        self.proj = nn.Conv2d(
            in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
        )
        if layer_norm:
            self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6)
        else:
            self.norm = None

        self.use_pos_embed = use_pos_embed
        if self.use_pos_embed:
            pos_embed = get_2d_sincos_pos_embed(embed_dim, int(num_patches**0.5))
            self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)

    def forward(self, latent):
        latent = self.proj(latent)
        if self.flatten:
            latent = latent.flatten(2).transpose(1, 2)  # BCHW -> BNC
        if self.layer_norm:
            latent = self.norm(latent)
        if self.use_pos_embed:
            return latent + self.pos_embed
        else:
            return latent


class SkipBlock(nn.Module):
    def __init__(self, dim: int):
        super().__init__()

        self.skip_linear = nn.Linear(2 * dim, dim)

        # Use torch.nn.LayerNorm for now, following the original code
        self.norm = nn.LayerNorm(dim)

    def forward(self, x, skip):
        x = self.skip_linear(torch.cat([x, skip], dim=-1))
        x = self.norm(x)

        return x


# Modified to support both pre-LayerNorm and post-LayerNorm configurations
# Don't support AdaLayerNormZero for now
# Modified from diffusers.models.attention.BasicTransformerBlock
class UTransformerBlock(nn.Module):
    r"""
    A modification of BasicTransformerBlock which supports pre-LayerNorm and post-LayerNorm configurations.

    Parameters:
        dim (`int`): The number of channels in the input and output.
        num_attention_heads (`int`): The number of heads to use for multi-head attention.
        attention_head_dim (`int`): The number of channels in each head.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
        activation_fn (`str`, *optional*, defaults to `"geglu"`):
            Activation function to be used in feed-forward.
        num_embeds_ada_norm (:obj: `int`, *optional*):
            The number of diffusion steps used during training. See `Transformer2DModel`.
        attention_bias (:obj: `bool`, *optional*, defaults to `False`):
            Configure if the attentions should contain a bias parameter.
        only_cross_attention (`bool`, *optional*):
            Whether to use only cross-attention layers. In this case two cross attention layers are used.
        double_self_attention (`bool`, *optional*):
            Whether to use two self-attention layers. In this case no cross attention layers are used.
        upcast_attention (`bool`, *optional*):
            Whether to upcast the query and key to float32 when performing the attention calculation.
        norm_elementwise_affine (`bool`, *optional*):
            Whether to use learnable per-element affine parameters during layer normalization.
        norm_type (`str`, defaults to `"layer_norm"`):
            The layer norm implementation to use.
        pre_layer_norm (`bool`, *optional*):
            Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
            as opposed to after ("post-LayerNorm"). Note that `BasicTransformerBlock` uses pre-LayerNorm, e.g.
            `pre_layer_norm = True`.
        final_dropout (`bool`, *optional*):
            Whether to use a final Dropout layer after the feedforward network.
    """

    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        dropout=0.0,
        cross_attention_dim: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        attention_bias: bool = False,
        only_cross_attention: bool = False,
        double_self_attention: bool = False,
        upcast_attention: bool = False,
        norm_elementwise_affine: bool = True,
        norm_type: str = "layer_norm",
        pre_layer_norm: bool = True,
        final_dropout: bool = False,
    ):
        super().__init__()
        self.only_cross_attention = only_cross_attention

        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"

        self.pre_layer_norm = pre_layer_norm

        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
            raise ValueError(
                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
            )

        # 1. Self-Attn
        self.attn1 = Attention(
            query_dim=dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
            upcast_attention=upcast_attention,
        )

        # 2. Cross-Attn
        if cross_attention_dim is not None or double_self_attention:
            self.attn2 = Attention(
                query_dim=dim,
                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                upcast_attention=upcast_attention,
            )  # is self-attn if encoder_hidden_states is none
        else:
            self.attn2 = None

        if self.use_ada_layer_norm:
            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
        else:
            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)

        if cross_attention_dim is not None or double_self_attention:
            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
            # the second cross attention block.
            self.norm2 = (
                AdaLayerNorm(dim, num_embeds_ada_norm)
                if self.use_ada_layer_norm
                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
            )
        else:
            self.norm2 = None

        # 3. Feed-forward
        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        timestep=None,
        cross_attention_kwargs=None,
        class_labels=None,
    ):
        # Pre-LayerNorm
        if self.pre_layer_norm:
            if self.use_ada_layer_norm:
                norm_hidden_states = self.norm1(hidden_states, timestep)
            else:
                norm_hidden_states = self.norm1(hidden_states)
        else:
            norm_hidden_states = hidden_states

        # 1. Self-Attention
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
        attn_output = self.attn1(
            norm_hidden_states,
            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
            attention_mask=attention_mask,
            **cross_attention_kwargs,
        )

        # Post-LayerNorm
        if not self.pre_layer_norm:
            if self.use_ada_layer_norm:
                attn_output = self.norm1(attn_output, timestep)
            else:
                attn_output = self.norm1(attn_output)

        hidden_states = attn_output + hidden_states

        if self.attn2 is not None:
            # Pre-LayerNorm
            if self.pre_layer_norm:
                norm_hidden_states = (
                    self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
                )
            else:
                norm_hidden_states = hidden_states
            # TODO (Birch-San): Here we should prepare the encoder_attention mask correctly
            # prepare attention mask here

            # 2. Cross-Attention
            attn_output = self.attn2(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                **cross_attention_kwargs,
            )

            # Post-LayerNorm
            if not self.pre_layer_norm:
                attn_output = self.norm2(attn_output, timestep) if self.use_ada_layer_norm else self.norm2(attn_output)

            hidden_states = attn_output + hidden_states

        # 3. Feed-forward
        # Pre-LayerNorm
        if self.pre_layer_norm:
            norm_hidden_states = self.norm3(hidden_states)
        else:
            norm_hidden_states = hidden_states

        ff_output = self.ff(norm_hidden_states)

        # Post-LayerNorm
        if not self.pre_layer_norm:
            ff_output = self.norm3(ff_output)

        hidden_states = ff_output + hidden_states

        return hidden_states


# Like UTransformerBlock except with LayerNorms on the residual backbone of the block
# Modified from diffusers.models.attention.BasicTransformerBlock
class UniDiffuserBlock(nn.Module):
    r"""
    A modification of BasicTransformerBlock which supports pre-LayerNorm and post-LayerNorm configurations and puts the
    LayerNorms on the residual backbone of the block. This matches the transformer block in the [original UniDiffuser
    implementation](https://github.com/thu-ml/unidiffuser/blob/main/libs/uvit_multi_post_ln_v1.py#L104).

    Parameters:
        dim (`int`): The number of channels in the input and output.
        num_attention_heads (`int`): The number of heads to use for multi-head attention.
        attention_head_dim (`int`): The number of channels in each head.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
        activation_fn (`str`, *optional*, defaults to `"geglu"`):
            Activation function to be used in feed-forward.
        num_embeds_ada_norm (:obj: `int`, *optional*):
            The number of diffusion steps used during training. See `Transformer2DModel`.
        attention_bias (:obj: `bool`, *optional*, defaults to `False`):
            Configure if the attentions should contain a bias parameter.
        only_cross_attention (`bool`, *optional*):
            Whether to use only cross-attention layers. In this case two cross attention layers are used.
        double_self_attention (`bool`, *optional*):
            Whether to use two self-attention layers. In this case no cross attention layers are used.
        upcast_attention (`bool`, *optional*):
            Whether to upcast the query and key to float() when performing the attention calculation.
        norm_elementwise_affine (`bool`, *optional*):
            Whether to use learnable per-element affine parameters during layer normalization.
        norm_type (`str`, defaults to `"layer_norm"`):
            The layer norm implementation to use.
        pre_layer_norm (`bool`, *optional*):
            Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
            as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm
            (`pre_layer_norm = False`).
        final_dropout (`bool`, *optional*):
            Whether to use a final Dropout layer after the feedforward network.
    """

    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        dropout=0.0,
        cross_attention_dim: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        attention_bias: bool = False,
        only_cross_attention: bool = False,
        double_self_attention: bool = False,
        upcast_attention: bool = False,
        norm_elementwise_affine: bool = True,
        norm_type: str = "layer_norm",
        pre_layer_norm: bool = False,
        final_dropout: bool = True,
    ):
        super().__init__()
        self.only_cross_attention = only_cross_attention

        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"

        self.pre_layer_norm = pre_layer_norm

        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
            raise ValueError(
                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
            )

        # 1. Self-Attn
        self.attn1 = Attention(
            query_dim=dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
            upcast_attention=upcast_attention,
        )

        # 2. Cross-Attn
        if cross_attention_dim is not None or double_self_attention:
            self.attn2 = Attention(
                query_dim=dim,
                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                upcast_attention=upcast_attention,
            )  # is self-attn if encoder_hidden_states is none
        else:
            self.attn2 = None

        if self.use_ada_layer_norm:
            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
        else:
            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)

        if cross_attention_dim is not None or double_self_attention:
            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
            # the second cross attention block.
            self.norm2 = (
                AdaLayerNorm(dim, num_embeds_ada_norm)
                if self.use_ada_layer_norm
                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
            )
        else:
            self.norm2 = None

        # 3. Feed-forward
        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        timestep=None,
        cross_attention_kwargs=None,
        class_labels=None,
    ):
        # Following the diffusers transformer block implementation, put the LayerNorm on the
        # residual backbone
        # Pre-LayerNorm
        if self.pre_layer_norm:
            if self.use_ada_layer_norm:
                hidden_states = self.norm1(hidden_states, timestep)
            else:
                hidden_states = self.norm1(hidden_states)

        # 1. Self-Attention
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
        attn_output = self.attn1(
            hidden_states,
            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
            attention_mask=attention_mask,
            **cross_attention_kwargs,
        )

        hidden_states = attn_output + hidden_states

        # Following the diffusers transformer block implementation, put the LayerNorm on the
        # residual backbone
        # Post-LayerNorm
        if not self.pre_layer_norm:
            if self.use_ada_layer_norm:
                hidden_states = self.norm1(hidden_states, timestep)
            else:
                hidden_states = self.norm1(hidden_states)

        if self.attn2 is not None:
            # Pre-LayerNorm
            if self.pre_layer_norm:
                hidden_states = (
                    self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
                )
            # TODO (Birch-San): Here we should prepare the encoder_attention mask correctly
            # prepare attention mask here

            # 2. Cross-Attention
            attn_output = self.attn2(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                **cross_attention_kwargs,
            )

            hidden_states = attn_output + hidden_states

            # Post-LayerNorm
            if not self.pre_layer_norm:
                hidden_states = (
                    self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
                )

        # 3. Feed-forward
        # Pre-LayerNorm
        if self.pre_layer_norm:
            hidden_states = self.norm3(hidden_states)

        ff_output = self.ff(hidden_states)

        hidden_states = ff_output + hidden_states

        # Post-LayerNorm
        if not self.pre_layer_norm:
            hidden_states = self.norm3(hidden_states)

        return hidden_states


# Modified from diffusers.models.transformer_2d.Transformer2DModel
# Modify the transformer block structure to be U-Net like following U-ViT
# Only supports patch-style input and torch.nn.LayerNorm currently
# https://github.com/baofff/U-ViT
class UTransformer2DModel(ModelMixin, ConfigMixin):
    """
    Transformer model based on the [U-ViT](https://github.com/baofff/U-ViT) architecture for image-like data. Compared
    to [`Transformer2DModel`], this model has skip connections between transformer blocks in a "U"-shaped fashion,
    similar to a U-Net. Supports only continuous (actual embeddings) inputs, which are embedded via a [`PatchEmbed`]
    layer and then reshaped to (b, t, d).

    Parameters:
        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
        in_channels (`int`, *optional*):
            Pass if the input is continuous. The number of channels in the input.
        out_channels (`int`, *optional*):
            The number of output channels; if `None`, defaults to `in_channels`.
        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        norm_num_groups (`int`, *optional*, defaults to `32`):
            The number of groups to use when performing Group Normalization.
        cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
        attention_bias (`bool`, *optional*):
            Configure if the TransformerBlocks' attention should contain a bias parameter.
        sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
            Note that this is fixed at training time as it is used for learning a number of position embeddings. See
            `ImagePositionalEmbeddings`.
        num_vector_embeds (`int`, *optional*):
            Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
            Includes the class for the masked latent pixel.
        patch_size (`int`, *optional*, defaults to 2):
            The patch size to use in the patch embedding.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
        num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
            The number of diffusion steps used during training. Note that this is fixed at training time as it is used
            to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
            up to but not more than steps than `num_embeds_ada_norm`.
        use_linear_projection (int, *optional*): TODO: Not used
        only_cross_attention (`bool`, *optional*):
            Whether to use only cross-attention layers. In this case two cross attention layers are used in each
            transformer block.
        upcast_attention (`bool`, *optional*):
            Whether to upcast the query and key to float() when performing the attention calculation.
        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
            The Layer Normalization implementation to use. Defaults to `torch.nn.LayerNorm`.
        block_type (`str`, *optional*, defaults to `"unidiffuser"`):
            The transformer block implementation to use. If `"unidiffuser"`, has the LayerNorms on the residual
            backbone of each transformer block; otherwise has them in the attention/feedforward branches (the standard
            behavior in `diffusers`.)
        pre_layer_norm (`bool`, *optional*):
            Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
            as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm
            (`pre_layer_norm = False`).
        norm_elementwise_affine (`bool`, *optional*):
            Whether to use learnable per-element affine parameters during layer normalization.
        use_patch_pos_embed (`bool`, *optional*):
            Whether to use position embeddings inside the patch embedding layer (`PatchEmbed`).
        final_dropout (`bool`, *optional*):
            Whether to use a final Dropout layer after the feedforward network.
    """

    @register_to_config
    def __init__(
        self,
        num_attention_heads: int = 16,
        attention_head_dim: int = 88,
        in_channels: Optional[int] = None,
        out_channels: Optional[int] = None,
        num_layers: int = 1,
        dropout: float = 0.0,
        norm_num_groups: int = 32,
        cross_attention_dim: Optional[int] = None,
        attention_bias: bool = False,
        sample_size: Optional[int] = None,
        num_vector_embeds: Optional[int] = None,
        patch_size: Optional[int] = 2,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        use_linear_projection: bool = False,
        only_cross_attention: bool = False,
        upcast_attention: bool = False,
        norm_type: str = "layer_norm",
        block_type: str = "unidiffuser",
        pre_layer_norm: bool = False,
        norm_elementwise_affine: bool = True,
        use_patch_pos_embed=False,
        ff_final_dropout: bool = False,
    ):
        super().__init__()
        self.use_linear_projection = use_linear_projection
        self.num_attention_heads = num_attention_heads
        self.attention_head_dim = attention_head_dim
        inner_dim = num_attention_heads * attention_head_dim

        # 1. Input
        # Only support patch input of shape (batch_size, num_channels, height, width) for now
        assert in_channels is not None and patch_size is not None, "Patch input requires in_channels and patch_size."

        assert sample_size is not None, "UTransformer2DModel over patched input must provide sample_size"

        # 2. Define input layers
        self.height = sample_size
        self.width = sample_size

        self.patch_size = patch_size
        self.pos_embed = PatchEmbed(
            height=sample_size,
            width=sample_size,
            patch_size=patch_size,
            in_channels=in_channels,
            embed_dim=inner_dim,
            use_pos_embed=use_patch_pos_embed,
        )

        # 3. Define transformers blocks
        # Modify this to have in_blocks ("downsample" blocks, even though we don't actually downsample), a mid_block,
        # and out_blocks ("upsample" blocks). Like a U-Net, there are skip connections from in_blocks to out_blocks in
        # a "U"-shaped fashion (e.g. first in_block to last out_block, etc.).
        # Quick hack to make the transformer block type configurable
        if block_type == "unidiffuser":
            block_cls = UniDiffuserBlock
        else:
            block_cls = UTransformerBlock
        self.transformer_in_blocks = nn.ModuleList(
            [
                block_cls(
                    inner_dim,
                    num_attention_heads,
                    attention_head_dim,
                    dropout=dropout,
                    cross_attention_dim=cross_attention_dim,
                    activation_fn=activation_fn,
                    num_embeds_ada_norm=num_embeds_ada_norm,
                    attention_bias=attention_bias,
                    only_cross_attention=only_cross_attention,
                    upcast_attention=upcast_attention,
                    norm_type=norm_type,
                    pre_layer_norm=pre_layer_norm,
                    norm_elementwise_affine=norm_elementwise_affine,
                    final_dropout=ff_final_dropout,
                )
                for d in range(num_layers // 2)
            ]
        )

        self.transformer_mid_block = block_cls(
            inner_dim,
            num_attention_heads,
            attention_head_dim,
            dropout=dropout,
            cross_attention_dim=cross_attention_dim,
            activation_fn=activation_fn,
            num_embeds_ada_norm=num_embeds_ada_norm,
            attention_bias=attention_bias,
            only_cross_attention=only_cross_attention,
            upcast_attention=upcast_attention,
            norm_type=norm_type,
            pre_layer_norm=pre_layer_norm,
            norm_elementwise_affine=norm_elementwise_affine,
            final_dropout=ff_final_dropout,
        )

        # For each skip connection, we use a SkipBlock (concatenation + Linear + LayerNorm) to process the inputs
        # before each transformer out_block.
        self.transformer_out_blocks = nn.ModuleList(
            [
                nn.ModuleDict(
                    {
                        "skip": SkipBlock(
                            inner_dim,
                        ),
                        "block": block_cls(
                            inner_dim,
                            num_attention_heads,
                            attention_head_dim,
                            dropout=dropout,
                            cross_attention_dim=cross_attention_dim,
                            activation_fn=activation_fn,
                            num_embeds_ada_norm=num_embeds_ada_norm,
                            attention_bias=attention_bias,
                            only_cross_attention=only_cross_attention,
                            upcast_attention=upcast_attention,
                            norm_type=norm_type,
                            pre_layer_norm=pre_layer_norm,
                            norm_elementwise_affine=norm_elementwise_affine,
                            final_dropout=ff_final_dropout,
                        ),
                    }
                )
                for d in range(num_layers // 2)
            ]
        )

        # 4. Define output layers
        self.out_channels = in_channels if out_channels is None else out_channels

        # Following the UniDiffuser U-ViT implementation, we process the transformer output with
        # a LayerNorm layer with per-element affine params
        self.norm_out = nn.LayerNorm(inner_dim)

    def forward(
        self,
        hidden_states,
        encoder_hidden_states=None,
        timestep=None,
        class_labels=None,
        cross_attention_kwargs=None,
        return_dict: bool = True,
        hidden_states_is_embedding: bool = False,
        unpatchify: bool = True,
    ):
        """
        Args:
            hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
                When continuous, `torch.Tensor` of shape `(batch size, channel, height, width)`): Input hidden_states
            encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
                self-attention.
            timestep ( `torch.long`, *optional*):
                Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
                Optional class labels to be applied as an embedding in AdaLayerZeroNorm. Used to indicate class labels
                conditioning.
            cross_attention_kwargs (*optional*):
                Keyword arguments to supply to the cross attention layers, if used.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
                tuple.
            hidden_states_is_embedding (`bool`, *optional*, defaults to `False`):
                Whether or not hidden_states is an embedding directly usable by the transformer. In this case we will
                ignore input handling (e.g. continuous, vectorized, etc.) and directly feed hidden_states into the
                transformer blocks.
            unpatchify (`bool`, *optional*, defaults to `True`):
                Whether to unpatchify the transformer output.

        Returns:
            [`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`:
            [`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When
            returning a tuple, the first element is the sample tensor.
        """
        # 0. Check inputs

        if not unpatchify and return_dict:
            raise ValueError(
                f"Cannot both define `unpatchify`: {unpatchify} and `return_dict`: {return_dict} since when"
                f" `unpatchify` is {unpatchify} the returned output is of shape (batch_size, seq_len, hidden_dim)"
                " rather than (batch_size, num_channels, height, width)."
            )

        # 1. Input
        if not hidden_states_is_embedding:
            hidden_states = self.pos_embed(hidden_states)

        # 2. Blocks

        # In ("downsample") blocks
        skips = []
        for in_block in self.transformer_in_blocks:
            hidden_states = in_block(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                timestep=timestep,
                cross_attention_kwargs=cross_attention_kwargs,
                class_labels=class_labels,
            )
            skips.append(hidden_states)

        # Mid block
        hidden_states = self.transformer_mid_block(hidden_states)

        # Out ("upsample") blocks
        for out_block in self.transformer_out_blocks:
            hidden_states = out_block["skip"](hidden_states, skips.pop())
            hidden_states = out_block["block"](
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                timestep=timestep,
                cross_attention_kwargs=cross_attention_kwargs,
                class_labels=class_labels,
            )

        # 3. Output
        # Don't support AdaLayerNorm for now, so no conditioning/scale/shift logic
        hidden_states = self.norm_out(hidden_states)
        # hidden_states = self.proj_out(hidden_states)

        if unpatchify:
            # unpatchify
            height = width = int(hidden_states.shape[1] ** 0.5)
            hidden_states = hidden_states.reshape(
                shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
            )
            hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
            output = hidden_states.reshape(
                shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
            )
        else:
            output = hidden_states

        if not return_dict:
            return (output,)

        return Transformer2DModelOutput(sample=output)


class UniDiffuserModel(ModelMixin, ConfigMixin):
    """
    Transformer model for a image-text [UniDiffuser](https://arxiv.org/pdf/2303.06555.pdf) model. This is a
    modification of [`UTransformer2DModel`] with input and output heads for the VAE-embedded latent image, the
    CLIP-embedded image, and the CLIP-embedded prompt (see paper for more details).

    Parameters:
        text_dim (`int`): The hidden dimension of the CLIP text model used to embed images.
        clip_img_dim (`int`): The hidden dimension of the CLIP vision model used to embed prompts.
        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
        in_channels (`int`, *optional*):
            Pass if the input is continuous. The number of channels in the input.
        out_channels (`int`, *optional*):
            The number of output channels; if `None`, defaults to `in_channels`.
        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        norm_num_groups (`int`, *optional*, defaults to `32`):
            The number of groups to use when performing Group Normalization.
        cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
        attention_bias (`bool`, *optional*):
            Configure if the TransformerBlocks' attention should contain a bias parameter.
        sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
            Note that this is fixed at training time as it is used for learning a number of position embeddings. See
            `ImagePositionalEmbeddings`.
        num_vector_embeds (`int`, *optional*):
            Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
            Includes the class for the masked latent pixel.
        patch_size (`int`, *optional*, defaults to 2):
            The patch size to use in the patch embedding.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
        num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
            The number of diffusion steps used during training. Note that this is fixed at training time as it is used
            to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
            up to but not more than steps than `num_embeds_ada_norm`.
        use_linear_projection (int, *optional*): TODO: Not used
        only_cross_attention (`bool`, *optional*):
            Whether to use only cross-attention layers. In this case two cross attention layers are used in each
            transformer block.
        upcast_attention (`bool`, *optional*):
            Whether to upcast the query and key to float32 when performing the attention calculation.
        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
            The Layer Normalization implementation to use. Defaults to `torch.nn.LayerNorm`.
        block_type (`str`, *optional*, defaults to `"unidiffuser"`):
            The transformer block implementation to use. If `"unidiffuser"`, has the LayerNorms on the residual
            backbone of each transformer block; otherwise has them in the attention/feedforward branches (the standard
            behavior in `diffusers`.)
        pre_layer_norm (`bool`, *optional*):
            Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
            as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm
            (`pre_layer_norm = False`).
        norm_elementwise_affine (`bool`, *optional*):
            Whether to use learnable per-element affine parameters during layer normalization.
        use_patch_pos_embed (`bool`, *optional*):
            Whether to use position embeddings inside the patch embedding layer (`PatchEmbed`).
        ff_final_dropout (`bool`, *optional*):
            Whether to use a final Dropout layer after the feedforward network.
        use_data_type_embedding (`bool`, *optional*):
            Whether to use a data type embedding. This is only relevant for UniDiffuser-v1 style models; UniDiffuser-v1
            is continue-trained from UniDiffuser-v0 on non-publically-available data and accepts a `data_type`
            argument, which can either be `1` to use the weights trained on non-publically-available data or `0`
            otherwise. This argument is subsequently embedded by the data type embedding, if used.
    """

    @register_to_config
    def __init__(
        self,
        text_dim: int = 768,
        clip_img_dim: int = 512,
        num_text_tokens: int = 77,
        num_attention_heads: int = 16,
        attention_head_dim: int = 88,
        in_channels: Optional[int] = None,
        out_channels: Optional[int] = None,
        num_layers: int = 1,
        dropout: float = 0.0,
        norm_num_groups: int = 32,
        cross_attention_dim: Optional[int] = None,
        attention_bias: bool = False,
        sample_size: Optional[int] = None,
        num_vector_embeds: Optional[int] = None,
        patch_size: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        use_linear_projection: bool = False,
        only_cross_attention: bool = False,
        upcast_attention: bool = False,
        norm_type: str = "layer_norm",
        block_type: str = "unidiffuser",
        pre_layer_norm: bool = False,
        use_timestep_embedding=False,
        norm_elementwise_affine: bool = True,
        use_patch_pos_embed=False,
        ff_final_dropout: bool = True,
        use_data_type_embedding: bool = False,
    ):
        super().__init__()

        # 0. Handle dimensions
        self.inner_dim = num_attention_heads * attention_head_dim

        assert sample_size is not None, "UniDiffuserModel over patched input must provide sample_size"
        self.sample_size = sample_size
        self.in_channels = in_channels
        self.out_channels = in_channels if out_channels is None else out_channels

        self.patch_size = patch_size
        # Assume image is square...
        self.num_patches = (self.sample_size // patch_size) * (self.sample_size // patch_size)

        # 1. Define input layers
        # 1.1 Input layers for text and image input
        # For now, only support patch input for VAE latent image input
        self.vae_img_in = PatchEmbed(
            height=sample_size,
            width=sample_size,
            patch_size=patch_size,
            in_channels=in_channels,
            embed_dim=self.inner_dim,
            use_pos_embed=use_patch_pos_embed,
        )
        self.clip_img_in = nn.Linear(clip_img_dim, self.inner_dim)
        self.text_in = nn.Linear(text_dim, self.inner_dim)

        # 1.2. Timestep embeddings for t_img, t_text
        self.timestep_img_proj = Timesteps(
            self.inner_dim,
            flip_sin_to_cos=True,
            downscale_freq_shift=0,
        )
        self.timestep_img_embed = (
            TimestepEmbedding(
                self.inner_dim,
                4 * self.inner_dim,
                out_dim=self.inner_dim,
            )
            if use_timestep_embedding
            else nn.Identity()
        )

        self.timestep_text_proj = Timesteps(
            self.inner_dim,
            flip_sin_to_cos=True,
            downscale_freq_shift=0,
        )
        self.timestep_text_embed = (
            TimestepEmbedding(
                self.inner_dim,
                4 * self.inner_dim,
                out_dim=self.inner_dim,
            )
            if use_timestep_embedding
            else nn.Identity()
        )

        # 1.3. Positional embedding
        self.num_text_tokens = num_text_tokens
        self.num_tokens = 1 + 1 + num_text_tokens + 1 + self.num_patches
        self.pos_embed = nn.Parameter(torch.zeros(1, self.num_tokens, self.inner_dim))
        self.pos_embed_drop = nn.Dropout(p=dropout)
        trunc_normal_(self.pos_embed, std=0.02)

        # 1.4. Handle data type token embeddings for UniDiffuser-V1, if necessary
        self.use_data_type_embedding = use_data_type_embedding
        if self.use_data_type_embedding:
            self.data_type_token_embedding = nn.Embedding(2, self.inner_dim)
            self.data_type_pos_embed_token = nn.Parameter(torch.zeros(1, 1, self.inner_dim))

        # 2. Define transformer blocks
        self.transformer = UTransformer2DModel(
            num_attention_heads=num_attention_heads,
            attention_head_dim=attention_head_dim,
            in_channels=in_channels,
            out_channels=out_channels,
            num_layers=num_layers,
            dropout=dropout,
            norm_num_groups=norm_num_groups,
            cross_attention_dim=cross_attention_dim,
            attention_bias=attention_bias,
            sample_size=sample_size,
            num_vector_embeds=num_vector_embeds,
            patch_size=patch_size,
            activation_fn=activation_fn,
            num_embeds_ada_norm=num_embeds_ada_norm,
            use_linear_projection=use_linear_projection,
            only_cross_attention=only_cross_attention,
            upcast_attention=upcast_attention,
            norm_type=norm_type,
            block_type=block_type,
            pre_layer_norm=pre_layer_norm,
            norm_elementwise_affine=norm_elementwise_affine,
            use_patch_pos_embed=use_patch_pos_embed,
            ff_final_dropout=ff_final_dropout,
        )

        # 3. Define output layers
        patch_dim = (patch_size**2) * out_channels
        self.vae_img_out = nn.Linear(self.inner_dim, patch_dim)
        self.clip_img_out = nn.Linear(self.inner_dim, clip_img_dim)
        self.text_out = nn.Linear(self.inner_dim, text_dim)

    @torch.jit.ignore
    def no_weight_decay(self):
        return {"pos_embed"}

    def forward(
        self,
        latent_image_embeds: torch.Tensor,
        image_embeds: torch.Tensor,
        prompt_embeds: torch.Tensor,
        timestep_img: Union[torch.Tensor, float, int],
        timestep_text: Union[torch.Tensor, float, int],
        data_type: Optional[Union[torch.Tensor, float, int]] = 1,
        encoder_hidden_states=None,
        cross_attention_kwargs=None,
    ):
        """
        Args:
            latent_image_embeds (`torch.Tensor` of shape `(batch size, latent channels, height, width)`):
                Latent image representation from the VAE encoder.
            image_embeds (`torch.Tensor` of shape `(batch size, 1, clip_img_dim)`):
                CLIP-embedded image representation (unsqueezed in the first dimension).
            prompt_embeds (`torch.Tensor` of shape `(batch size, seq_len, text_dim)`):
                CLIP-embedded text representation.
            timestep_img (`torch.long` or `float` or `int`):
                Current denoising step for the image.
            timestep_text (`torch.long` or `float` or `int`):
                Current denoising step for the text.
            data_type: (`torch.int` or `float` or `int`, *optional*, defaults to `1`):
                Only used in UniDiffuser-v1-style models. Can be either `1`, to use weights trained on nonpublic data,
                or `0` otherwise.
            encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
                self-attention.
            cross_attention_kwargs (*optional*):
                Keyword arguments to supply to the cross attention layers, if used.


        Returns:
            `tuple`: Returns relevant parts of the model's noise prediction: the first element of the tuple is tbe VAE
            image embedding, the second element is the CLIP image embedding, and the third element is the CLIP text
            embedding.
        """
        batch_size = latent_image_embeds.shape[0]

        # 1. Input
        # 1.1. Map inputs to shape (B, N, inner_dim)
        vae_hidden_states = self.vae_img_in(latent_image_embeds)
        clip_hidden_states = self.clip_img_in(image_embeds)
        text_hidden_states = self.text_in(prompt_embeds)

        num_text_tokens, num_img_tokens = text_hidden_states.size(1), vae_hidden_states.size(1)

        # 1.2. Encode image timesteps to single token (B, 1, inner_dim)
        if not torch.is_tensor(timestep_img):
            timestep_img = torch.tensor([timestep_img], dtype=torch.long, device=vae_hidden_states.device)

        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
        timestep_img = timestep_img * torch.ones(batch_size, dtype=timestep_img.dtype, device=timestep_img.device)

        timestep_img_token = self.timestep_img_proj(timestep_img)
        # t_img_token does not contain any weights and will always return f32 tensors
        # but time_embedding might be fp16, so we need to cast here.
        timestep_img_token = timestep_img_token.to(dtype=self.dtype)
        timestep_img_token = self.timestep_img_embed(timestep_img_token)
        timestep_img_token = timestep_img_token.unsqueeze(dim=1)

        # 1.3. Encode text timesteps to single token (B, 1, inner_dim)
        if not torch.is_tensor(timestep_text):
            timestep_text = torch.tensor([timestep_text], dtype=torch.long, device=vae_hidden_states.device)

        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
        timestep_text = timestep_text * torch.ones(batch_size, dtype=timestep_text.dtype, device=timestep_text.device)

        timestep_text_token = self.timestep_text_proj(timestep_text)
        # t_text_token does not contain any weights and will always return f32 tensors
        # but time_embedding might be fp16, so we need to cast here.
        timestep_text_token = timestep_text_token.to(dtype=self.dtype)
        timestep_text_token = self.timestep_text_embed(timestep_text_token)
        timestep_text_token = timestep_text_token.unsqueeze(dim=1)

        # 1.4. Concatenate all of the embeddings together.
        if self.use_data_type_embedding:
            assert data_type is not None, "data_type must be supplied if the model uses a data type embedding"
            if not torch.is_tensor(data_type):
                data_type = torch.tensor([data_type], dtype=torch.int, device=vae_hidden_states.device)

            # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
            data_type = data_type * torch.ones(batch_size, dtype=data_type.dtype, device=data_type.device)

            data_type_token = self.data_type_token_embedding(data_type).unsqueeze(dim=1)
            hidden_states = torch.cat(
                [
                    timestep_img_token,
                    timestep_text_token,
                    data_type_token,
                    text_hidden_states,
                    clip_hidden_states,
                    vae_hidden_states,
                ],
                dim=1,
            )
        else:
            hidden_states = torch.cat(
                [timestep_img_token, timestep_text_token, text_hidden_states, clip_hidden_states, vae_hidden_states],
                dim=1,
            )

        # 1.5. Prepare the positional embeddings and add to hidden states
        # Note: I think img_vae should always have the proper shape, so there's no need to interpolate
        # the position embeddings.
        if self.use_data_type_embedding:
            pos_embed = torch.cat(
                [self.pos_embed[:, : 1 + 1, :], self.data_type_pos_embed_token, self.pos_embed[:, 1 + 1 :, :]], dim=1
            )
        else:
            pos_embed = self.pos_embed
        hidden_states = hidden_states + pos_embed
        hidden_states = self.pos_embed_drop(hidden_states)

        # 2. Blocks
        hidden_states = self.transformer(
            hidden_states,
            encoder_hidden_states=encoder_hidden_states,
            timestep=None,
            class_labels=None,
            cross_attention_kwargs=cross_attention_kwargs,
            return_dict=False,
            hidden_states_is_embedding=True,
            unpatchify=False,
        )[0]

        # 3. Output
        # Split out the predicted noise representation.
        if self.use_data_type_embedding:
            (
                t_img_token_out,
                t_text_token_out,
                data_type_token_out,
                text_out,
                img_clip_out,
                img_vae_out,
            ) = hidden_states.split((1, 1, 1, num_text_tokens, 1, num_img_tokens), dim=1)
        else:
            t_img_token_out, t_text_token_out, text_out, img_clip_out, img_vae_out = hidden_states.split(
                (1, 1, num_text_tokens, 1, num_img_tokens), dim=1
            )

        img_vae_out = self.vae_img_out(img_vae_out)

        # unpatchify
        height = width = int(img_vae_out.shape[1] ** 0.5)
        img_vae_out = img_vae_out.reshape(
            shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
        )
        img_vae_out = torch.einsum("nhwpqc->nchpwq", img_vae_out)
        img_vae_out = img_vae_out.reshape(
            shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
        )

        img_clip_out = self.clip_img_out(img_clip_out)

        text_out = self.text_out(text_out)

        return img_vae_out, img_clip_out, text_out