Delete controlnet_union.py

controlnet_union.py

@@ -1,1061 +0,0 @@
-from collections import OrderedDict
-from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, Tuple, Union
-
-import torch
-from diffusers.configuration_utils import ConfigMixin, register_to_config
-from diffusers.loaders import FromOriginalModelMixin
-from diffusers.models.attention_processor import (
-    ADDED_KV_ATTENTION_PROCESSORS,
-    CROSS_ATTENTION_PROCESSORS,
-    AttentionProcessor,
-    AttnAddedKVProcessor,
-    AttnProcessor,
-)
-from diffusers.models.embeddings import (
-    TextImageProjection,
-    TextImageTimeEmbedding,
-    TextTimeEmbedding,
-    TimestepEmbedding,
-    Timesteps,
-)
-from diffusers.models.modeling_utils import ModelMixin
-from diffusers.models.unets.unet_2d_blocks import (
-    CrossAttnDownBlock2D,
-    DownBlock2D,
-    UNetMidBlock2DCrossAttn,
-    get_down_block,
-)
-from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
-from diffusers.utils import BaseOutput, logging
-from torch import nn
-from torch.nn import functional as F
-
-logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
-
-
-# Transformer Block
-# Used to exchange info between different conditions and input image
-# With reference to https://github.com/TencentARC/T2I-Adapter/blob/SD/ldm/modules/encoders/adapter.py#L147
-class QuickGELU(nn.Module):
-    def forward(self, x: torch.Tensor):
-        return x * torch.sigmoid(1.702 * x)
-
-
-class LayerNorm(nn.LayerNorm):
-    """Subclass torch's LayerNorm to handle fp16."""
-
-    def forward(self, x: torch.Tensor):
-        orig_type = x.dtype
-        ret = super().forward(x)
-        return ret.type(orig_type)
-
-
-class ResidualAttentionBlock(nn.Module):
-    def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
-        super().__init__()
-
-        self.attn = nn.MultiheadAttention(d_model, n_head)
-        self.ln_1 = LayerNorm(d_model)
-        self.mlp = nn.Sequential(
-            OrderedDict(
-                [
-                    ("c_fc", nn.Linear(d_model, d_model * 4)),
-                    ("gelu", QuickGELU()),
-                    ("c_proj", nn.Linear(d_model * 4, d_model)),
-                ]
-            )
-        )
-        self.ln_2 = LayerNorm(d_model)
-        self.attn_mask = attn_mask
-
-    def attention(self, x: torch.Tensor):
-        self.attn_mask = (
-            self.attn_mask.to(dtype=x.dtype, device=x.device)
-            if self.attn_mask is not None
-            else None
-        )
-        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
-
-    def forward(self, x: torch.Tensor):
-        x = x + self.attention(self.ln_1(x))
-        x = x + self.mlp(self.ln_2(x))
-        return x
-
-
-# -----------------------------------------------------------------------------------------------------
-
-@dataclass
-class ControlNetOutput(BaseOutput):
-    """
-    The output of [`ControlNetModel`].
-    Args:
-        down_block_res_samples (`tuple[torch.Tensor]`):
-            A tuple of downsample activations at different resolutions for each downsampling block. Each tensor should
-            be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
-            used to condition the original UNet's downsampling activations.
-        mid_down_block_re_sample (`torch.Tensor`):
-            The activation of the midde block (the lowest sample resolution). Each tensor should be of shape
-            `(batch_size, channel * lowest_resolution, height // lowest_resolution, width // lowest_resolution)`.
-            Output can be used to condition the original UNet's middle block activation.
-    """
-
-    down_block_res_samples: Tuple[torch.Tensor]
-    mid_block_res_sample: torch.Tensor
-
-
-class ControlNetConditioningEmbedding(nn.Module):
-    """
-    Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
-    [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
-    training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
-    convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
-    (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
-    model) to encode image-space conditions ... into feature maps ..."
-    """
-
-    # original setting is (16, 32, 96, 256)
-    def __init__(
-        self,
-        conditioning_embedding_channels: int,
-        conditioning_channels: int = 3,
-        block_out_channels: Tuple[int] = (48, 96, 192, 384),
-    ):
-        super().__init__()
-
-        self.conv_in = nn.Conv2d(
-            conditioning_channels, block_out_channels[0], kernel_size=3, padding=1
-        )
-
-        self.blocks = nn.ModuleList([])
-
-        for i in range(len(block_out_channels) - 1):
-            channel_in = block_out_channels[i]
-            channel_out = block_out_channels[i + 1]
-            self.blocks.append(
-                nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1)
-            )
-            self.blocks.append(
-                nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=2)
-            )
-
-        self.conv_out = zero_module(
-            nn.Conv2d(
-                block_out_channels[-1],
-                conditioning_embedding_channels,
-                kernel_size=3,
-                padding=1,
-            )
-        )
-
-    def forward(self, conditioning):
-        embedding = self.conv_in(conditioning)
-        embedding = F.silu(embedding)
-
-        for block in self.blocks:
-            embedding = block(embedding)
-            embedding = F.silu(embedding)
-
-        embedding = self.conv_out(embedding)
-
-        return embedding
-
-
-class ControlNetModel_Union(ModelMixin, ConfigMixin, FromOriginalModelMixin):
-    """
-    A ControlNet model.
-    Args:
-        in_channels (`int`, defaults to 4):
-            The number of channels in the input sample.
-        flip_sin_to_cos (`bool`, defaults to `True`):
-            Whether to flip the sin to cos in the time embedding.
-        freq_shift (`int`, defaults to 0):
-            The frequency shift to apply to the time embedding.
-        down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
-            The tuple of downsample blocks to use.
-        only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
-        block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
-            The tuple of output channels for each block.
-        layers_per_block (`int`, defaults to 2):
-            The number of layers per block.
-        downsample_padding (`int`, defaults to 1):
-            The padding to use for the downsampling convolution.
-        mid_block_scale_factor (`float`, defaults to 1):
-            The scale factor to use for the mid block.
-        act_fn (`str`, 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, normalization and activation layers is skipped
-            in post-processing.
-        norm_eps (`float`, defaults to 1e-5):
-            The epsilon to use for the normalization.
-        cross_attention_dim (`int`, defaults to 1280):
-            The dimension of the cross attention features.
-        transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
-            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
-            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
-            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
-        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 are down-projected to text
-            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
-        attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8):
-            The dimension of the attention heads.
-        use_linear_projection (`bool`, defaults to `False`):
-        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 0):
-            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`.
-        upcast_attention (`bool`, defaults to `False`):
-        resnet_time_scale_shift (`str`, defaults to `"default"`):
-            Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
-        projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
-            The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
-            `class_embed_type="projection"`.
-        controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
-            The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
-        conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
-            The tuple of output channel for each block in the `conditioning_embedding` layer.
-        global_pool_conditions (`bool`, defaults to `False`):
-    """
-
-    _supports_gradient_checkpointing = True
-
-    @register_to_config
-    def __init__(
-        self,
-        in_channels: int = 4,
-        conditioning_channels: int = 3,
-        flip_sin_to_cos: bool = True,
-        freq_shift: int = 0,
-        down_block_types: Tuple[str] = (
-            "CrossAttnDownBlock2D",
-            "CrossAttnDownBlock2D",
-            "CrossAttnDownBlock2D",
-            "DownBlock2D",
-        ),
-        only_cross_attention: Union[bool, Tuple[bool]] = False,
-        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
-        layers_per_block: 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: int = 1280,
-        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
-        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,
-        use_linear_projection: bool = False,
-        class_embed_type: Optional[str] = None,
-        addition_embed_type: Optional[str] = None,
-        addition_time_embed_dim: Optional[int] = None,
-        num_class_embeds: Optional[int] = None,
-        upcast_attention: bool = False,
-        resnet_time_scale_shift: str = "default",
-        projection_class_embeddings_input_dim: Optional[int] = None,
-        controlnet_conditioning_channel_order: str = "rgb",
-        conditioning_embedding_out_channels: Optional[Tuple[int]] = (16, 32, 96, 256),
-        global_pool_conditions: bool = False,
-        addition_embed_type_num_heads=64,
-        num_control_type=6,
-    ):
-        super().__init__()
-
-        # 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(block_out_channels) != len(down_block_types):
-            raise ValueError(
-                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
-            )
-
-        if not isinstance(only_cross_attention, bool) and len(
-            only_cross_attention
-        ) != len(down_block_types):
-            raise ValueError(
-                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
-            )
-
-        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(
-            down_block_types
-        ):
-            raise ValueError(
-                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
-            )
-
-        if isinstance(transformer_layers_per_block, int):
-            transformer_layers_per_block = [transformer_layers_per_block] * len(
-                down_block_types
-            )
-
-        # input
-        conv_in_kernel = 3
-        conv_in_padding = (conv_in_kernel - 1) // 2
-        self.conv_in = nn.Conv2d(
-            in_channels,
-            block_out_channels[0],
-            kernel_size=conv_in_kernel,
-            padding=conv_in_padding,
-        )
-
-        # time
-        time_embed_dim = block_out_channels[0] * 4
-        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
-        timestep_input_dim = block_out_channels[0]
-        self.time_embedding = TimestepEmbedding(
-            timestep_input_dim,
-            time_embed_dim,
-            act_fn=act_fn,
-        )
-
-        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)
-        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
-            )
-        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 == "text_time":
-            self.add_time_proj = Timesteps(
-                addition_time_embed_dim, flip_sin_to_cos, freq_shift
-            )
-            self.add_embedding = TimestepEmbedding(
-                projection_class_embeddings_input_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'."
-            )
-
-        # control net conditioning embedding
-        self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
-            conditioning_embedding_channels=block_out_channels[0],
-            block_out_channels=conditioning_embedding_out_channels,
-            conditioning_channels=conditioning_channels,
-        )
-
-        # Copyright by Qi Xin(2024/07/06)
-        # Condition Transformer(fuse single/multi conditions with input image)
-        # The Condition Transformer augment the feature representation of conditions
-        # The overall design is somewhat like resnet. The output of Condition Transformer is used to predict a condition bias adding to the original condition feature.
-        # num_control_type = 6
-        num_trans_channel = 320
-        num_trans_head = 8
-        num_trans_layer = 1
-        num_proj_channel = 320
-        task_scale_factor = num_trans_channel**0.5
-
-        self.task_embedding = nn.Parameter(
-            task_scale_factor * torch.randn(num_control_type, num_trans_channel)
-        )
-        self.transformer_layes = nn.Sequential(
-            *[
-                ResidualAttentionBlock(num_trans_channel, num_trans_head)
-                for _ in range(num_trans_layer)
-            ]
-        )
-        self.spatial_ch_projs = zero_module(
-            nn.Linear(num_trans_channel, num_proj_channel)
-        )
-        # -----------------------------------------------------------------------------------------------------
-
-        # Copyright by Qi Xin(2024/07/06)
-        # Control Encoder to distinguish different control conditions
-        # A simple but effective module, consists of an embedding layer and a linear layer, to inject the control info to time embedding.
-        self.control_type_proj = Timesteps(
-            addition_time_embed_dim, flip_sin_to_cos, freq_shift
-        )
-        self.control_add_embedding = TimestepEmbedding(
-            addition_time_embed_dim * num_control_type, time_embed_dim
-        )
-        # -----------------------------------------------------------------------------------------------------
-
-        self.down_blocks = nn.ModuleList([])
-        self.controlnet_down_blocks = nn.ModuleList([])
-
-        if isinstance(only_cross_attention, bool):
-            only_cross_attention = [only_cross_attention] * len(down_block_types)
-
-        if isinstance(attention_head_dim, int):
-            attention_head_dim = (attention_head_dim,) * len(down_block_types)
-
-        if isinstance(num_attention_heads, int):
-            num_attention_heads = (num_attention_heads,) * len(down_block_types)
-
-        # down
-        output_channel = block_out_channels[0]
-
-        controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
-        controlnet_block = zero_module(controlnet_block)
-        self.controlnet_down_blocks.append(controlnet_block)
-
-        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,
-                transformer_layers_per_block=transformer_layers_per_block[i],
-                in_channels=input_channel,
-                out_channels=output_channel,
-                temb_channels=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,
-                num_attention_heads=num_attention_heads[i],
-                attention_head_dim=attention_head_dim[i]
-                if attention_head_dim[i] is not None
-                else output_channel,
-                downsample_padding=downsample_padding,
-                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,
-            )
-            self.down_blocks.append(down_block)
-
-            for _ in range(layers_per_block):
-                controlnet_block = nn.Conv2d(
-                    output_channel, output_channel, kernel_size=1
-                )
-                controlnet_block = zero_module(controlnet_block)
-                self.controlnet_down_blocks.append(controlnet_block)
-
-            if not is_final_block:
-                controlnet_block = nn.Conv2d(
-                    output_channel, output_channel, kernel_size=1
-                )
-                controlnet_block = zero_module(controlnet_block)
-                self.controlnet_down_blocks.append(controlnet_block)
-
-        # mid
-        mid_block_channel = block_out_channels[-1]
-
-        controlnet_block = nn.Conv2d(
-            mid_block_channel, mid_block_channel, kernel_size=1
-        )
-        controlnet_block = zero_module(controlnet_block)
-        self.controlnet_mid_block = controlnet_block
-
-        self.mid_block = UNetMidBlock2DCrossAttn(
-            transformer_layers_per_block=transformer_layers_per_block[-1],
-            in_channels=mid_block_channel,
-            temb_channels=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,
-            num_attention_heads=num_attention_heads[-1],
-            resnet_groups=norm_num_groups,
-            use_linear_projection=use_linear_projection,
-            upcast_attention=upcast_attention,
-        )
-
-    @classmethod
-    def from_unet(
-        cls,
-        unet: UNet2DConditionModel,
-        controlnet_conditioning_channel_order: str = "rgb",
-        conditioning_embedding_out_channels: Optional[Tuple[int]] = (16, 32, 96, 256),
-        load_weights_from_unet: bool = True,
-    ):
-        r"""
-        Instantiate a [`ControlNetModel`] from [`UNet2DConditionModel`].
-        Parameters:
-            unet (`UNet2DConditionModel`):
-                The UNet model weights to copy to the [`ControlNetModel`]. All configuration options are also copied
-                where applicable.
-        """
-        transformer_layers_per_block = (
-            unet.config.transformer_layers_per_block
-            if "transformer_layers_per_block" in unet.config
-            else 1
-        )
-        encoder_hid_dim = (
-            unet.config.encoder_hid_dim if "encoder_hid_dim" in unet.config else None
-        )
-        encoder_hid_dim_type = (
-            unet.config.encoder_hid_dim_type
-            if "encoder_hid_dim_type" in unet.config
-            else None
-        )
-        addition_embed_type = (
-            unet.config.addition_embed_type
-            if "addition_embed_type" in unet.config
-            else None
-        )
-        addition_time_embed_dim = (
-            unet.config.addition_time_embed_dim
-            if "addition_time_embed_dim" in unet.config
-            else None
-        )
-
-        controlnet = cls(
-            encoder_hid_dim=encoder_hid_dim,
-            encoder_hid_dim_type=encoder_hid_dim_type,
-            addition_embed_type=addition_embed_type,
-            addition_time_embed_dim=addition_time_embed_dim,
-            transformer_layers_per_block=transformer_layers_per_block,
-            # transformer_layers_per_block=[1, 2, 5],
-            in_channels=unet.config.in_channels,
-            flip_sin_to_cos=unet.config.flip_sin_to_cos,
-            freq_shift=unet.config.freq_shift,
-            down_block_types=unet.config.down_block_types,
-            only_cross_attention=unet.config.only_cross_attention,
-            block_out_channels=unet.config.block_out_channels,
-            layers_per_block=unet.config.layers_per_block,
-            downsample_padding=unet.config.downsample_padding,
-            mid_block_scale_factor=unet.config.mid_block_scale_factor,
-            act_fn=unet.config.act_fn,
-            norm_num_groups=unet.config.norm_num_groups,
-            norm_eps=unet.config.norm_eps,
-            cross_attention_dim=unet.config.cross_attention_dim,
-            attention_head_dim=unet.config.attention_head_dim,
-            num_attention_heads=unet.config.num_attention_heads,
-            use_linear_projection=unet.config.use_linear_projection,
-            class_embed_type=unet.config.class_embed_type,
-            num_class_embeds=unet.config.num_class_embeds,
-            upcast_attention=unet.config.upcast_attention,
-            resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
-            projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim,
-            controlnet_conditioning_channel_order=controlnet_conditioning_channel_order,
-            conditioning_embedding_out_channels=conditioning_embedding_out_channels,
-        )
-
-        if load_weights_from_unet:
-            controlnet.conv_in.load_state_dict(unet.conv_in.state_dict())
-            controlnet.time_proj.load_state_dict(unet.time_proj.state_dict())
-            controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
-
-            if controlnet.class_embedding:
-                controlnet.class_embedding.load_state_dict(
-                    unet.class_embedding.state_dict()
-                )
-
-            controlnet.down_blocks.load_state_dict(
-                unet.down_blocks.state_dict(), strict=False
-            )
-            controlnet.mid_block.load_state_dict(
-                unet.mid_block.state_dict(), strict=False
-            )
-
-        return controlnet
-
-    @property
-    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(
-            name: str,
-            module: torch.nn.Module,
-            processors: Dict[str, AttentionProcessor],
-        ):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor(
-                    return_deprecated_lora=True
-                )
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(
-        self,
-        processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]],
-        _remove_lora=False,
-    ):
-        r"""
-        Sets the attention processor to use to compute attention.
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
-            if hasattr(module, "set_processor"):
-                if not isinstance(processor, dict):
-                    module.set_processor(processor, _remove_lora=_remove_lora)
-                else:
-                    module.set_processor(
-                        processor.pop(f"{name}.processor"), _remove_lora=_remove_lora
-                    )
-
-            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)
-
-    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
-    def set_default_attn_processor(self):
-        """
-        Disables custom attention processors and sets the default attention implementation.
-        """
-        if all(
-            proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS
-            for proc in self.attn_processors.values()
-        ):
-            processor = AttnAddedKVProcessor()
-        elif all(
-            proc.__class__ in CROSS_ATTENTION_PROCESSORS
-            for proc in self.attn_processors.values()
-        ):
-            processor = AttnProcessor()
-        else:
-            raise ValueError(
-                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
-            )
-
-        self.set_attn_processor(processor, _remove_lora=True)
-
-    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attention_slice
-    def set_attention_slice(self, slice_size):
-        r"""
-        Enable sliced attention computation.
-        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
-        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
-        Args:
-            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
-                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
-                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
-                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
-                must be a multiple of `slice_size`.
-        """
-        sliceable_head_dims = []
-
-        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
-            if hasattr(module, "set_attention_slice"):
-                sliceable_head_dims.append(module.sliceable_head_dim)
-
-            for child in module.children():
-                fn_recursive_retrieve_sliceable_dims(child)
-
-        # retrieve number of attention layers
-        for module in self.children():
-            fn_recursive_retrieve_sliceable_dims(module)
-
-        num_sliceable_layers = len(sliceable_head_dims)
-
-        if slice_size == "auto":
-            # half the attention head size is usually a good trade-off between
-            # speed and memory
-            slice_size = [dim // 2 for dim in sliceable_head_dims]
-        elif slice_size == "max":
-            # make smallest slice possible
-            slice_size = num_sliceable_layers * [1]
-
-        slice_size = (
-            num_sliceable_layers * [slice_size]
-            if not isinstance(slice_size, list)
-            else slice_size
-        )
-
-        if len(slice_size) != len(sliceable_head_dims):
-            raise ValueError(
-                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
-                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
-            )
-
-        for i in range(len(slice_size)):
-            size = slice_size[i]
-            dim = sliceable_head_dims[i]
-            if size is not None and size > dim:
-                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
-
-        # Recursively walk through all the children.
-        # Any children which exposes the set_attention_slice method
-        # gets the message
-        def fn_recursive_set_attention_slice(
-            module: torch.nn.Module, slice_size: List[int]
-        ):
-            if hasattr(module, "set_attention_slice"):
-                module.set_attention_slice(slice_size.pop())
-
-            for child in module.children():
-                fn_recursive_set_attention_slice(child, slice_size)
-
-        reversed_slice_size = list(reversed(slice_size))
-        for module in self.children():
-            fn_recursive_set_attention_slice(module, reversed_slice_size)
-
-    def _set_gradient_checkpointing(self, module, value=False):
-        if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
-            module.gradient_checkpointing = value
-
-    def forward(
-        self,
-        sample: torch.FloatTensor,
-        timestep: Union[torch.Tensor, float, int],
-        encoder_hidden_states: torch.Tensor,
-        controlnet_cond_list: torch.FloatTensor,
-        conditioning_scale: float = 1.0,
-        class_labels: Optional[torch.Tensor] = None,
-        timestep_cond: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
-        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
-        guess_mode: bool = False,
-        return_dict: bool = True,
-    ) -> Union[ControlNetOutput, Tuple]:
-        """
-        The [`ControlNetModel`] forward method.
-        Args:
-            sample (`torch.FloatTensor`):
-                The noisy input tensor.
-            timestep (`Union[torch.Tensor, float, int]`):
-                The number of timesteps to denoise an input.
-            encoder_hidden_states (`torch.Tensor`):
-                The encoder hidden states.
-            controlnet_cond (`torch.FloatTensor`):
-                The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
-            conditioning_scale (`float`, defaults to `1.0`):
-                The scale factor for ControlNet outputs.
-            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
-                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
-            timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
-                Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
-                timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
-                embeddings.
-            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
-                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
-                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
-                negative values to the attention scores corresponding to "discard" tokens.
-            added_cond_kwargs (`dict`):
-                Additional conditions for the Stable Diffusion XL UNet.
-            cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
-                A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
-            guess_mode (`bool`, defaults to `False`):
-                In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
-                you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
-            return_dict (`bool`, defaults to `True`):
-                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
-        Returns:
-            [`~models.controlnet.ControlNetOutput`] **or** `tuple`:
-                If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
-                returned where the first element is the sample tensor.
-        """
-        # check channel order
-        channel_order = self.config.controlnet_conditioning_channel_order
-
-        if channel_order == "rgb":
-            # in rgb order by default
-            ...
-        # elif channel_order == "bgr":
-        #     controlnet_cond = torch.flip(controlnet_cond, dims=[1])
-        else:
-            raise ValueError(
-                f"unknown `controlnet_conditioning_channel_order`: {channel_order}"
-            )
-
-        # prepare attention_mask
-        if attention_mask is not None:
-            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
-            attention_mask = attention_mask.unsqueeze(1)
-
-        # 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)
-        aug_emb = None
-
-        if self.class_embedding is not None:
-            if class_labels is None:
-                raise ValueError(
-                    "class_labels should be provided when num_class_embeds > 0"
-                )
-
-            if self.config.class_embed_type == "timestep":
-                class_labels = self.time_proj(class_labels)
-
-            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
-            emb = emb + class_emb
-
-        if self.config.addition_embed_type is not None:
-            if self.config.addition_embed_type == "text":
-                aug_emb = self.add_embedding(encoder_hidden_states)
-
-            elif self.config.addition_embed_type == "text_time":
-                if "text_embeds" not in added_cond_kwargs:
-                    raise ValueError(
-                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
-                    )
-                text_embeds = added_cond_kwargs.get("text_embeds")
-                if "time_ids" not in added_cond_kwargs:
-                    raise ValueError(
-                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
-                    )
-                time_ids = added_cond_kwargs.get("time_ids")
-                time_embeds = self.add_time_proj(time_ids.flatten())
-                time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
-
-                add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
-                add_embeds = add_embeds.to(emb.dtype)
-                aug_emb = self.add_embedding(add_embeds)
-
-        # Copyright by Qi Xin(2024/07/06)
-        # inject control type info to time embedding to distinguish different control conditions
-        control_type = added_cond_kwargs.get("control_type")
-        control_embeds = self.control_type_proj(control_type.flatten())
-        control_embeds = control_embeds.reshape((t_emb.shape[0], -1))
-        control_embeds = control_embeds.to(emb.dtype)
-        control_emb = self.control_add_embedding(control_embeds)
-        emb = emb + control_emb
-        # ---------------------------------------------------------------------------------
-
-        emb = emb + aug_emb if aug_emb is not None else emb
-
-        # 2. pre-process
-        sample = self.conv_in(sample)
-        indices = torch.nonzero(control_type[0])
-
-        # Copyright by Qi Xin(2024/07/06)
-        # add single/multi conditons to input image.
-        # Condition Transformer provides an easy and effective way to fuse different features naturally
-        inputs = []
-        condition_list = []
-
-        for idx in range(indices.shape[0] + 1):
-            if idx == indices.shape[0]:
-                controlnet_cond = sample
-                feat_seq = torch.mean(controlnet_cond, dim=(2, 3))  # N * C
-            else:
-                controlnet_cond = self.controlnet_cond_embedding(
-                    controlnet_cond_list[indices[idx][0]]
-                )
-                feat_seq = torch.mean(controlnet_cond, dim=(2, 3))  # N * C
-                feat_seq = feat_seq + self.task_embedding[indices[idx][0]]
-
-            inputs.append(feat_seq.unsqueeze(1))
-            condition_list.append(controlnet_cond)
-
-        x = torch.cat(inputs, dim=1)  # NxLxC
-        x = self.transformer_layes(x)
-
-        controlnet_cond_fuser = sample * 0.0
-        for idx in range(indices.shape[0]):
-            alpha = self.spatial_ch_projs(x[:, idx])
-            alpha = alpha.unsqueeze(-1).unsqueeze(-1)
-            controlnet_cond_fuser += condition_list[idx] + alpha
-
-        sample = sample + controlnet_cond_fuser
-        # -------------------------------------------------------------------------------------------
-
-        # 3. down
-        down_block_res_samples = (sample,)
-        for downsample_block in self.down_blocks:
-            if (
-                hasattr(downsample_block, "has_cross_attention")
-                and downsample_block.has_cross_attention
-            ):
-                sample, res_samples = downsample_block(
-                    hidden_states=sample,
-                    temb=emb,
-                    encoder_hidden_states=encoder_hidden_states,
-                    attention_mask=attention_mask,
-                    cross_attention_kwargs=cross_attention_kwargs,
-                )
-            else:
-                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
-
-            down_block_res_samples += res_samples
-
-        # 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,
-            )
-
-        # 5. Control net blocks
-
-        controlnet_down_block_res_samples = ()
-
-        for down_block_res_sample, controlnet_block in zip(
-            down_block_res_samples, self.controlnet_down_blocks
-        ):
-            down_block_res_sample = controlnet_block(down_block_res_sample)
-            controlnet_down_block_res_samples = controlnet_down_block_res_samples + (
-                down_block_res_sample,
-            )
-
-        down_block_res_samples = controlnet_down_block_res_samples
-
-        mid_block_res_sample = self.controlnet_mid_block(sample)
-
-        # 6. scaling
-        if guess_mode and not self.config.global_pool_conditions:
-            scales = torch.logspace(
-                -1, 0, len(down_block_res_samples) + 1, device=sample.device
-            )  # 0.1 to 1.0
-            scales = scales * conditioning_scale
-            down_block_res_samples = [
-                sample * scale for sample, scale in zip(down_block_res_samples, scales)
-            ]
-            mid_block_res_sample = mid_block_res_sample * scales[-1]  # last one
-        else:
-            down_block_res_samples = [
-                sample * conditioning_scale for sample in down_block_res_samples
-            ]
-            mid_block_res_sample = mid_block_res_sample * conditioning_scale
-
-        if self.config.global_pool_conditions:
-            down_block_res_samples = [
-                torch.mean(sample, dim=(2, 3), keepdim=True)
-                for sample in down_block_res_samples
-            ]
-            mid_block_res_sample = torch.mean(
-                mid_block_res_sample, dim=(2, 3), keepdim=True
-            )
-
-        if not return_dict:
-            return (down_block_res_samples, mid_block_res_sample)
-
-        return ControlNetOutput(
-            down_block_res_samples=down_block_res_samples,
-            mid_block_res_sample=mid_block_res_sample,
-        )
-
-
-def zero_module(module):
-    for p in module.parameters():
-        nn.init.zeros_(p)
-    return module