mvdiffusion/models/unet_mv2d_condition.py

# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple, Union
import os

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

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders import UNet2DConditionLoadersMixin
from diffusers.utils import BaseOutput, logging
from diffusers.models.activations import get_activation
from diffusers.models.attention_processor import AttentionProcessor, AttnProcessor
from diffusers.models.embeddings import (
    GaussianFourierProjection,
    ImageHintTimeEmbedding,
    ImageProjection,
    ImageTimeEmbedding,
    TextImageProjection,
    TextImageTimeEmbedding,
    TextTimeEmbedding,
    TimestepEmbedding,
    Timesteps,
)
from diffusers.models.modeling_utils import ModelMixin, load_state_dict
from diffusers.models.unets.unet_2d_blocks import (
    CrossAttnDownBlock2D,
    CrossAttnUpBlock2D,
    DownBlock2D,
    UNetMidBlock2DCrossAttn,
    UNetMidBlock2DSimpleCrossAttn,
    UpBlock2D,
)
from diffusers.utils import (
    CONFIG_NAME,
    FLAX_WEIGHTS_NAME,
    SAFETENSORS_WEIGHTS_NAME,
    WEIGHTS_NAME,
    _add_variant,
    _get_model_file,
    deprecate,
    is_torch_version,
    logging,
)
from diffusers.utils.import_utils import is_accelerate_available 
from diffusers.utils.hub_utils import HF_HUB_OFFLINE
from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE
DIFFUSERS_CACHE = HUGGINGFACE_HUB_CACHE

from diffusers import __version__
from .unet_mv2d_blocks import (
    CrossAttnDownBlockMV2D,
    CrossAttnUpBlockMV2D,
    UNetMidBlockMV2DCrossAttn,
    get_down_block,
    get_up_block,
)
from einops import rearrange, repeat

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


@dataclass
class UNetMV2DConditionOutput(BaseOutput):
    """
    The output of [`UNet2DConditionModel`].

    Args:
        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
    """

    sample: torch.FloatTensor = None


class ResidualBlock(nn.Module):
    def __init__(self, dim):
        super(ResidualBlock, self).__init__()
        self.linear1 = nn.Linear(dim, dim)
        self.activation = nn.SiLU()
        self.linear2 = nn.Linear(dim, dim)

    def forward(self, x):
        identity = x
        out = self.linear1(x)
        out = self.activation(out)
        out = self.linear2(out)
        out += identity  
        out = self.activation(out)
        return out
    
class ResidualLiner(nn.Module):
    def __init__(self, in_features, out_features, dim, act=None, num_block=1):
        super(ResidualLiner, self).__init__()
        self.linear_in = nn.Sequential(nn.Linear(in_features, dim), nn.SiLU())
        
        blocks = nn.ModuleList()
        for _ in range(num_block):
            blocks.append(ResidualBlock(dim))
        self.blocks = blocks    
        
        self.linear_out = nn.Linear(dim, out_features)
        self.act = act

    def forward(self, x):
        out = self.linear_in(x)
        for block in self.blocks:
            out = block(out)
        out = self.linear_out(out)
        if self.act is not None:
            out = self.act(out)
        return out

class BasicConvBlock(nn.Module):
    def __init__(self, in_channels, out_channels, stride=1):
        super(BasicConvBlock, self).__init__()
        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1, bias=False)
        self.norm1 = nn.GroupNorm(num_groups=8, num_channels=in_channels,  affine=True)
        self.act = nn.SiLU()
        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)
        self.norm2 = nn.GroupNorm(num_groups=8, num_channels=in_channels,  affine=True)
        self.downsample = nn.Sequential()
        if stride != 1 or in_channels != out_channels:
            self.downsample = nn.Sequential(
                nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride, bias=False),
                nn.GroupNorm(num_groups=8, num_channels=in_channels,  affine=True)
            )

    def forward(self, x):
        identity = x
        out = self.conv1(x)
        out = self.norm1(out)
        out = self.act(out)
        out = self.conv2(out)
        out = self.norm2(out)
        out += self.downsample(identity)
        out = self.act(out)
        return out

class UNetMV2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
    r"""
    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
    shaped output.

    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
    for all models (such as downloading or saving).

    Parameters:
        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
            Height and width of input/output sample.
        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
            Whether to flip the sin to cos in the time embedding.
        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
            The tuple of downsample blocks to use.
        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
            Block type for middle of UNet, it can be either `UNetMidBlock2DCrossAttn` or
            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
            The tuple of upsample blocks to use.
        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
            Whether to include self-attention in the basic transformer blocks, see
            [`~models.attention.BasicTransformerBlock`].
        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
            The tuple of output channels for each block.
        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
            If `None`, normalization and activation layers is skipped in post-processing.
        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
            The dimension of the cross attention features.
        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 (`int`, *optional*, defaults to 8): The dimension of the attention heads.
        num_attention_heads (`int`, *optional*):
            The number of attention heads. If not defined, defaults to `attention_head_dim`
        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
        class_embed_type (`str`, *optional*, defaults to `None`):
            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
        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.
        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
            Dimension for the timestep embeddings.
        num_class_embeds (`int`, *optional*, defaults to `None`):
            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
            class conditioning with `class_embed_type` equal to `None`.
        time_embedding_type (`str`, *optional*, defaults to `positional`):
            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
        time_embedding_dim (`int`, *optional*, defaults to `None`):
            An optional override for the dimension of the projected time embedding.
        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
            Optional activation function to use only once on the time embeddings before they are passed to the rest of
            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
        timestep_post_act (`str`, *optional*, defaults to `None`):
            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
            The dimension of `cond_proj` layer in the timestep embedding.
        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
        conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
        projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
            embeddings with the class embeddings.
        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
            otherwise.
    """

    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        sample_size: Optional[int] = None,
        in_channels: int = 4,
        out_channels: int = 4,
        center_input_sample: bool = False,
        flip_sin_to_cos: bool = True,
        freq_shift: int = 0,
        down_block_types: Tuple[str] = (
            "CrossAttnDownBlockMV2D",
            "CrossAttnDownBlockMV2D",
            "CrossAttnDownBlockMV2D",
            "DownBlock2D",
        ),
        mid_block_type: Optional[str] = "UNetMidBlockMV2DCrossAttn",
        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlockMV2D", "CrossAttnUpBlockMV2D", "CrossAttnUpBlockMV2D"),
        only_cross_attention: Union[bool, Tuple[bool]] = False,
        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
        layers_per_block: Union[int, Tuple[int]] = 2,
        downsample_padding: int = 1,
        mid_block_scale_factor: float = 1,
        act_fn: str = "silu",
        norm_num_groups: Optional[int] = 32,
        norm_eps: float = 1e-5,
        cross_attention_dim: Union[int, Tuple[int]] = 1280,
        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,
        dual_cross_attention: bool = False,
        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",
        resnet_skip_time_act: bool = False,
        resnet_out_scale_factor: int = 1.0,
        time_embedding_type: str = "positional",
        time_embedding_dim: Optional[int] = None,
        time_embedding_act_fn: Optional[str] = None,
        timestep_post_act: Optional[str] = None,
        time_cond_proj_dim: Optional[int] = None,
        conv_in_kernel: int = 3,
        conv_out_kernel: int = 3,
        projection_class_embeddings_input_dim: Optional[int] = None,
        class_embeddings_concat: bool = False,
        mid_block_only_cross_attention: Optional[bool] = None,
        cross_attention_norm: Optional[str] = None,
        addition_embed_type_num_heads=64,
        num_views: int = 1,
        cd_attention_last: bool = False,
        cd_attention_mid: bool = False,
        multiview_attention: bool = True,
        sparse_mv_attention: bool = False,
        selfattn_block: str = "custom",
        addition_downsample: bool = False,
        addition_channels: Optional[Tuple[int]] = (1280, 1280, 1280),
    ):
        super().__init__()

        self.sample_size = sample_size
        self.num_views = num_views

        if num_attention_heads is not None:
            raise ValueError(
                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
            )

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

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

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

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

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

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

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

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

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

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

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

        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
            encoder_hid_dim_type = "text_proj"
            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")

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

        if encoder_hid_dim_type == "text_proj":
            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
        elif encoder_hid_dim_type == "text_image_proj":
            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
            self.encoder_hid_proj = TextImageProjection(
                text_embed_dim=encoder_hid_dim,
                image_embed_dim=cross_attention_dim,
                cross_attention_dim=cross_attention_dim,
            )
        elif encoder_hid_dim_type == "image_proj":
            # Kandinsky 2.2
            self.encoder_hid_proj = ImageProjection(
                image_embed_dim=encoder_hid_dim,
                cross_attention_dim=cross_attention_dim,
            )
        elif encoder_hid_dim_type is not None:
            raise ValueError(
                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
            )
        else:
            self.encoder_hid_proj = None

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

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

            self.add_embedding = TextTimeEmbedding(
                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
            )
        elif addition_embed_type == "text_image":
            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
            self.add_embedding = TextImageTimeEmbedding(
                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
            )
        elif addition_embed_type == "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 == "image":
            # Kandinsky 2.2
            self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
        elif addition_embed_type == "image_hint":
            # Kandinsky 2.2 ControlNet
            self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
        elif addition_embed_type is not None:
            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")

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

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

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

            only_cross_attention = [only_cross_attention] * len(down_block_types)

        if mid_block_only_cross_attention is None:
            mid_block_only_cross_attention = False

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

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

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

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

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

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

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

            down_block = get_down_block(
                down_block_type,
                num_layers=layers_per_block[i],
                transformer_layers_per_block=transformer_layers_per_block[i],
                in_channels=input_channel,
                out_channels=output_channel,
                temb_channels=blocks_time_embed_dim,
                add_downsample=not is_final_block,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                resnet_groups=norm_num_groups,
                cross_attention_dim=cross_attention_dim[i],
                num_attention_heads=num_attention_heads[i],
                downsample_padding=downsample_padding,
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                only_cross_attention=only_cross_attention[i],
                upcast_attention=upcast_attention,
                resnet_time_scale_shift=resnet_time_scale_shift,
                resnet_skip_time_act=resnet_skip_time_act,
                resnet_out_scale_factor=resnet_out_scale_factor,
                cross_attention_norm=cross_attention_norm,
                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
                num_views=num_views,
                cd_attention_last=cd_attention_last,
                cd_attention_mid=cd_attention_mid,
                multiview_attention=multiview_attention,
                sparse_mv_attention=sparse_mv_attention,
                selfattn_block=selfattn_block,
            )
            self.down_blocks.append(down_block)

        # mid
        if mid_block_type == "UNetMidBlock2DCrossAttn":
            self.mid_block = UNetMidBlock2DCrossAttn(
                transformer_layers_per_block=transformer_layers_per_block[-1],
                in_channels=block_out_channels[-1],
                temb_channels=blocks_time_embed_dim,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
                resnet_time_scale_shift=resnet_time_scale_shift,
                cross_attention_dim=cross_attention_dim[-1],
                num_attention_heads=num_attention_heads[-1],
                resnet_groups=norm_num_groups,
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                upcast_attention=upcast_attention,
            )
        # custom MV2D attention block  
        elif mid_block_type == "UNetMidBlockMV2DCrossAttn":
            self.mid_block = UNetMidBlockMV2DCrossAttn(
                transformer_layers_per_block=transformer_layers_per_block[-1],
                in_channels=block_out_channels[-1],
                temb_channels=blocks_time_embed_dim,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
                resnet_time_scale_shift=resnet_time_scale_shift,
                cross_attention_dim=cross_attention_dim[-1],
                num_attention_heads=num_attention_heads[-1],
                resnet_groups=norm_num_groups,
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                upcast_attention=upcast_attention,
                num_views=num_views,
                cd_attention_last=cd_attention_last,
                cd_attention_mid=cd_attention_mid,
                multiview_attention=multiview_attention,
                sparse_mv_attention=sparse_mv_attention,
                selfattn_block=selfattn_block,
            )
        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
                in_channels=block_out_channels[-1],
                temb_channels=blocks_time_embed_dim,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
                cross_attention_dim=cross_attention_dim[-1],
                attention_head_dim=attention_head_dim[-1],
                resnet_groups=norm_num_groups,
                resnet_time_scale_shift=resnet_time_scale_shift,
                skip_time_act=resnet_skip_time_act,
                only_cross_attention=mid_block_only_cross_attention,
                cross_attention_norm=cross_attention_norm,
            )
        elif mid_block_type is None:
            self.mid_block = None
        else:
            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
        
        self.addition_downsample = addition_downsample
        if self.addition_downsample:
            inc = block_out_channels[-1]
            self.downsample = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
            self.conv_block = nn.ModuleList()
            self.conv_block.append(BasicConvBlock(inc, addition_channels[0], stride=1)) 
            for dim_ in addition_channels[1:-1]:  
                self.conv_block.append(BasicConvBlock(dim_, dim_, stride=1))    
            self.conv_block.append(BasicConvBlock(dim_, inc))
            self.addition_conv_out = nn.Conv2d(inc, inc, kernel_size=1, bias=False)
            nn.init.zeros_(self.addition_conv_out.weight.data)
            self.addition_act_out = nn.SiLU()
            self.upsample = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)

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

        # up
        reversed_block_out_channels = list(reversed(block_out_channels))
        reversed_num_attention_heads = list(reversed(num_attention_heads))
        reversed_layers_per_block = list(reversed(layers_per_block))
        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
        reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
        only_cross_attention = list(reversed(only_cross_attention))

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

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

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

            up_block = get_up_block(
                up_block_type,
                num_layers=reversed_layers_per_block[i] + 1,
                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
                in_channels=input_channel,
                out_channels=output_channel,
                prev_output_channel=prev_output_channel,
                temb_channels=blocks_time_embed_dim,
                add_upsample=add_upsample,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                resnet_groups=norm_num_groups,
                cross_attention_dim=reversed_cross_attention_dim[i],
                num_attention_heads=reversed_num_attention_heads[i],
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                only_cross_attention=only_cross_attention[i],
                upcast_attention=upcast_attention,
                resnet_time_scale_shift=resnet_time_scale_shift,
                resnet_skip_time_act=resnet_skip_time_act,
                resnet_out_scale_factor=resnet_out_scale_factor,
                cross_attention_norm=cross_attention_norm,
                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
                num_views=num_views,
                cd_attention_last=cd_attention_last,
                cd_attention_mid=cd_attention_mid,
                multiview_attention=multiview_attention,
                sparse_mv_attention=sparse_mv_attention,
                selfattn_block=selfattn_block,
            )
            self.up_blocks.append(up_block)
            prev_output_channel = output_channel

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

            self.conv_act = get_activation(act_fn)

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

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

    @property
    def attn_processors(self) -> Dict[str, AttentionProcessor]:
        r"""
        Returns:
            `dict` of attention processors: A dictionary containing all attention processors used in the model with
            indexed by its weight name.
        """
        # set recursively
        processors = {}

        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
            if hasattr(module, "set_processor"):
                processors[f"{name}.processor"] = module.processor

            for sub_name, child in module.named_children():
                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

            return processors

        for name, module in self.named_children():
            fn_recursive_add_processors(name, module, processors)

        return processors

    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
        r"""
        Sets the attention processor to use to compute attention.

        Parameters:
            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
                The instantiated processor class or a dictionary of processor classes that will be set as the processor
                for **all** `Attention` layers.

                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
                processor. This is strongly recommended when setting trainable attention processors.

        """
        count = len(self.attn_processors.keys())

        if isinstance(processor, dict) and len(processor) != count:
            raise ValueError(
                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
            )

        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
            if hasattr(module, "set_processor"):
                if not isinstance(processor, dict):
                    module.set_processor(processor)
                else:
                    module.set_processor(processor.pop(f"{name}.processor"))

            for sub_name, child in module.named_children():
                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

        for name, module in self.named_children():
            fn_recursive_attn_processor(name, module, processor)

    def set_default_attn_processor(self):
        """
        Disables custom attention processors and sets the default attention implementation.
        """
        self.set_attn_processor(AttnProcessor())

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

        When this option is enabled, the attention module 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, CrossAttnDownBlockMV2D, DownBlock2D, CrossAttnUpBlock2D, CrossAttnUpBlockMV2D, UpBlock2D)):
            module.gradient_checkpointing = value

    def forward(
        self,
        sample: torch.FloatTensor,
        timestep: Union[torch.Tensor, float, int],
        encoder_hidden_states: torch.Tensor = None,
        class_labels: Optional[torch.Tensor] = None,
        timestep_cond: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
        mid_block_additional_residual: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        return_dict: bool = True,
        vis_max_min: bool = False,
    ) -> Union[UNetMV2DConditionOutput, Tuple]:
        r"""
        The [`UNet2DConditionModel`] forward method.

        Args:
            sample (`torch.FloatTensor`):
                The noisy input tensor with the following shape `(batch, channel, height, width)`.
            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
            encoder_hidden_states (`torch.FloatTensor`):
                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
            encoder_attention_mask (`torch.Tensor`):
                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
                which adds large negative values to the attention scores corresponding to "discard" tokens.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
                tuple.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
            added_cond_kwargs: (`dict`, *optional*):
                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
                are passed along to the UNet blocks.

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

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

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

        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
        # expects mask of shape:
        #   [batch, key_tokens]
        # adds singleton query_tokens dimension:
        #   [batch,                    1, key_tokens]
        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
        if attention_mask is not None:
            # assume that mask is expressed as:
            #   (1 = keep,      0 = discard)
            # convert mask into a bias that can be added to attention scores:
            #       (keep = +0,     discard = -10000.0)
            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
            attention_mask = attention_mask.unsqueeze(1)

        # convert encoder_attention_mask to a bias the same way we do for attention_mask
        if encoder_attention_mask is not None:
            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)

        # 0. center input if necessary
        if self.config.center_input_sample:
            sample = 2 * sample - 1.0
        # 1. time
        timesteps = timestep
        if not torch.is_tensor(timesteps):
            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
            # This would be a good case for the `match` statement (Python 3.10+)
            is_mps = sample.device.type == "mps"
            if isinstance(timestep, float):
                dtype = torch.float32 if is_mps else torch.float64
            else:
                dtype = torch.int32 if is_mps else torch.int64
            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
        elif len(timesteps.shape) == 0:
            timesteps = timesteps[None].to(sample.device)

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

        t_emb = self.time_proj(timesteps)

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

        emb = self.time_embedding(t_emb, timestep_cond)
        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)

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

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

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

            image_embs = added_cond_kwargs.get("image_embeds")
            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
            aug_emb = self.add_embedding(text_embs, image_embs)
        elif self.config.addition_embed_type == "text_time":
            # SDXL - style
            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)
        elif self.config.addition_embed_type == "image":
            # Kandinsky 2.2 - style
            if "image_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
                )
            image_embs = added_cond_kwargs.get("image_embeds")
            aug_emb = self.add_embedding(image_embs)
        elif self.config.addition_embed_type == "image_hint":
            # Kandinsky 2.2 - style
            if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
                )
            image_embs = added_cond_kwargs.get("image_embeds")
            hint = added_cond_kwargs.get("hint")
            aug_emb, hint = self.add_embedding(image_embs, hint)
            sample = torch.cat([sample, hint], dim=1)

        emb = emb + aug_emb if aug_emb is not None else emb
        emb_pre_act = emb
        if self.time_embed_act is not None:
            emb = self.time_embed_act(emb)

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

            image_embeds = added_cond_kwargs.get("image_embeds")
            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
            # Kandinsky 2.2 - style
            if "image_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
                )
            image_embeds = added_cond_kwargs.get("image_embeds")
            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
        # 2. pre-process
        sample = self.conv_in(sample)
        # 3. down

        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
        is_adapter = mid_block_additional_residual is None and down_block_additional_residuals is not None

        down_block_res_samples = (sample,)
        for i, downsample_block in enumerate(self.down_blocks):
            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
                # For t2i-adapter CrossAttnDownBlock2D
                additional_residuals = {}
                if is_adapter and len(down_block_additional_residuals) > 0:
                    additional_residuals["additional_residuals"] = down_block_additional_residuals.pop(0)

                sample, res_samples = downsample_block(
                    hidden_states=sample,
                    temb=emb,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                    encoder_attention_mask=encoder_attention_mask,
                    **additional_residuals,
                )
            else:
                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)

                if is_adapter and len(down_block_additional_residuals) > 0:
                    sample += down_block_additional_residuals.pop(0)

            down_block_res_samples += res_samples

        if is_controlnet:
            new_down_block_res_samples = ()

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

            down_block_res_samples = new_down_block_res_samples

        if self.addition_downsample:
            global_sample = sample
            global_sample = self.downsample(global_sample)
            for layer in self.conv_block:
                global_sample = layer(global_sample)
            global_sample = self.addition_act_out(self.addition_conv_out(global_sample))
            global_sample = self.upsample(global_sample)
        # 4. mid
        if self.mid_block is not None:
            sample = self.mid_block(
                sample,
                emb,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=attention_mask,
                cross_attention_kwargs=cross_attention_kwargs,
                encoder_attention_mask=encoder_attention_mask,
            )
            
        if is_controlnet:
            sample = sample + mid_block_additional_residual

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

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

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

            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
                sample = upsample_block(
                    hidden_states=sample,
                    temb=emb,
                    res_hidden_states_tuple=res_samples,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    upsample_size=upsample_size,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                )
            else:
                sample = upsample_block(
                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
                )
        if torch.isnan(sample).any() or torch.isinf(sample).any():
            print("NAN in sample, stop training.")
            exit() 
        # 6. post-process
        if self.conv_norm_out:
            sample = self.conv_norm_out(sample)
            sample = self.conv_act(sample)
        sample = self.conv_out(sample)
        if not return_dict:
            return sample
        return UNetMV2DConditionOutput(sample=sample)

        
    @classmethod
    def from_pretrained_2d(
            cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]],
            num_views: int, sample_size: int,
            zero_init_conv_in: bool = True,
            cd_attention_last: bool = False, 
            cd_attention_mid: bool = False, multiview_attention: bool = True, 
            sparse_mv_attention: bool = False, selfattn_block: str = 'custom',
            in_channels: int = 8, out_channels: int = 4, unclip: bool = False, 
            init_mvattn_with_selfattn: bool= False, addition_downsample: bool = False,
            **kwargs
        ):
        r"""
        Instantiate a pretrained PyTorch model from a pretrained model configuration.

        The model is set in evaluation mode - `model.eval()` - by default, and dropout modules are deactivated. To
        train the model, set it back in training mode with `model.train()`.

        Parameters:
            pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
                Can be either:

                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
                      the Hub.
                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
                      with [`~ModelMixin.save_pretrained`].

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            torch_dtype (`str` or `torch.dtype`, *optional*):
                Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
                dtype is automatically derived from the model's weights.
            force_download (`bool`, *optional*, defaults to `False`):
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.

            proxies (`Dict[str, str]`, *optional*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            output_loading_info (`bool`, *optional*, defaults to `False`):
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
            local_files_only(`bool`, *optional*, defaults to `False`):
                Whether to only load local model weights and configuration files or not. If set to `True`, the model
                won't be downloaded from the Hub.
            use_auth_token (`str` or *bool*, *optional*):
                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                `diffusers-cli login` (stored in `~/.huggingface`) is used.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            from_flax (`bool`, *optional*, defaults to `False`):
                Load the model weights from a Flax checkpoint save file.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
            mirror (`str`, *optional*):
                Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                information.
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
                A map that specifies where each submodule should go. It doesn't need to be defined for each
                parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
                same device.

                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
                more information about each option see [designing a device
                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
            max_memory (`Dict`, *optional*):
                A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
                each GPU and the available CPU RAM if unset.
            offload_folder (`str` or `os.PathLike`, *optional*):
                The path to offload weights if `device_map` contains the value `"disk"`.
            offload_state_dict (`bool`, *optional*):
                If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
                the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
                when there is some disk offload.
            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
                argument to `True` will raise an error.
            variant (`str`, *optional*):
                Load weights from a specified `variant` filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.
            use_safetensors (`bool`, *optional*, defaults to `None`):
                If set to `None`, the `safetensors` weights are downloaded if they're available **and** if the
                `safetensors` library is installed. If set to `True`, the model is forcibly loaded from `safetensors`
                weights. If set to `False`, `safetensors` weights are not loaded.

        <Tip>

        To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with
        `huggingface-cli login`. You can also activate the special
        ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use this method in a
        firewalled environment.

        </Tip>

        Example:

        ```py
        from diffusers import UNet2DConditionModel

        unet = UNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="unet")
        ```

        If you get the error message below, you need to finetune the weights for your downstream task:

        ```bash
        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
        - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
        You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
        ```
        """
        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
        ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
        force_download = kwargs.pop("force_download", False)
        from_flax = kwargs.pop("from_flax", False)
        resume_download = kwargs.pop("resume_download", False)
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
        use_auth_token = kwargs.pop("use_auth_token", None)
        revision = kwargs.pop("revision", None)
        torch_dtype = kwargs.pop("torch_dtype", None)
        subfolder = kwargs.pop("subfolder", None)
        device_map = kwargs.pop("device_map", None)
        max_memory = kwargs.pop("max_memory", None)
        offload_folder = kwargs.pop("offload_folder", None)
        offload_state_dict = kwargs.pop("offload_state_dict", False)
        variant = kwargs.pop("variant", None)
        use_safetensors = kwargs.pop("use_safetensors", None)

        if use_safetensors:
            raise ValueError(
                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetensors"
            )

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

        if device_map is not None and not is_accelerate_available():
            raise NotImplementedError(
                "Loading and dispatching requires `accelerate`. Please make sure to install accelerate or set"
                " `device_map=None`. You can install accelerate with `pip install accelerate`."
            )

        # Check if we can handle device_map and dispatching the weights
        if device_map is not None and not is_torch_version(">=", "1.9.0"):
            raise NotImplementedError(
                "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
                " `device_map=None`."
            )

        # Load config if we don't provide a configuration
        config_path = pretrained_model_name_or_path

        user_agent = {
            "diffusers": __version__,
            "file_type": "model",
            "framework": "pytorch",
        }

        # load config
        config, unused_kwargs, commit_hash = cls.load_config(
            config_path,
            cache_dir=cache_dir,
            return_unused_kwargs=True,
            return_commit_hash=True,
            force_download=force_download,
            resume_download=resume_download,
            proxies=proxies,
            local_files_only=local_files_only,
            use_auth_token=use_auth_token,
            revision=revision,
            subfolder=subfolder,
            device_map=device_map,
            max_memory=max_memory,
            offload_folder=offload_folder,
            offload_state_dict=offload_state_dict,
            user_agent=user_agent,
            **kwargs,
        )

        # modify config
        config["_class_name"] = cls.__name__
        config['in_channels'] = in_channels
        config['out_channels'] = out_channels
        config['sample_size'] = sample_size # training resolution
        config['num_views'] = num_views
        config['cd_attention_last'] = cd_attention_last
        config['cd_attention_mid'] = cd_attention_mid
        config['multiview_attention'] = multiview_attention
        config['sparse_mv_attention'] = sparse_mv_attention
        config['selfattn_block'] = selfattn_block
        config["down_block_types"] = [
            "CrossAttnDownBlockMV2D",
            "CrossAttnDownBlockMV2D",
            "CrossAttnDownBlockMV2D",
            "DownBlock2D"
        ]
        config['mid_block_type'] = "UNetMidBlockMV2DCrossAttn"
        config["up_block_types"] = [
            "UpBlock2D",
            "CrossAttnUpBlockMV2D",
            "CrossAttnUpBlockMV2D",
            "CrossAttnUpBlockMV2D"
        ]
        
        config['addition_downsample'] = addition_downsample
        # load model
        model_file = None
        if from_flax:
            raise NotImplementedError
        else:
            if use_safetensors:
                try:
                    model_file = _get_model_file(
                        pretrained_model_name_or_path,
                        weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant),
                        cache_dir=cache_dir,
                        force_download=force_download,
                        proxies=proxies,
                        local_files_only=local_files_only,
                        use_auth_token=use_auth_token,
                        revision=revision,
                        subfolder=subfolder,
                        user_agent=user_agent,
                        commit_hash=commit_hash,
                    )
                except IOError as e:
                    if not allow_pickle:
                        raise e
                    pass
            if model_file is None:
                model_file = _get_model_file(
                    pretrained_model_name_or_path,
                    weights_name=_add_variant(WEIGHTS_NAME, variant),
                    cache_dir=cache_dir,
                    force_download=force_download,
                    proxies=proxies,
                    local_files_only=local_files_only,
                    use_auth_token=use_auth_token,
                    revision=revision,
                    subfolder=subfolder,
                    user_agent=user_agent,
                    commit_hash=commit_hash,
                )

            model = cls.from_config(config, **unused_kwargs)
            import copy
            state_dict_pretrain = load_state_dict(model_file, variant=variant)
            state_dict = copy.deepcopy(state_dict_pretrain)
            
            if init_mvattn_with_selfattn:
                for key in state_dict_pretrain:
                    if 'attn1' in key:
                        key_mv = key.replace('attn1', 'attn_mv')
                        state_dict[key_mv] = state_dict_pretrain[key]
                        if 'to_out.0.weight' in key:
                            nn.init.zeros_(state_dict[key_mv].data)
                    if 'transformer_blocks' in key and 'norm1' in key: # in case that initialize the norm layer in resnet block
                        key_mv = key.replace('norm1', 'norm_mv')
                        state_dict[key_mv] = state_dict_pretrain[key]
            # del state_dict_pretrain
            
            model._convert_deprecated_attention_blocks(state_dict)

            conv_in_weight = state_dict['conv_in.weight']
            conv_out_weight = state_dict['conv_out.weight']
            model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model_2d(
                model,
                state_dict,
                model_file,
                pretrained_model_name_or_path,
                ignore_mismatched_sizes=True,
            )
            if any([key == 'conv_in.weight' for key, _, _ in mismatched_keys]):
                # initialize from the original SD structure
                model.conv_in.weight.data[:,:4] = conv_in_weight

            # whether to place all zero to new layers?
            if zero_init_conv_in:
                model.conv_in.weight.data[:,4:] = 0.

            if any([key == 'conv_out.weight' for key, _, _ in mismatched_keys]):
                # initialize from the original SD structure
                model.conv_out.weight.data[:,:4] = conv_out_weight
                if out_channels == 8: # copy for the last 4 channels
                    model.conv_out.weight.data[:, 4:] = conv_out_weight

            loading_info = {
                "missing_keys": missing_keys,
                "unexpected_keys": unexpected_keys,
                "mismatched_keys": mismatched_keys,
                "error_msgs": error_msgs,
            }

        if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
            raise ValueError(
                f"{torch_dtype} needs to be of type `torch.dtype`, e.g. `torch.float16`, but is {type(torch_dtype)}."
            )
        elif torch_dtype is not None:
            model = model.to(torch_dtype)

        model.register_to_config(_name_or_path=pretrained_model_name_or_path)

        # Set model in evaluation mode to deactivate DropOut modules by default
        model.eval()
        if output_loading_info:
            return model, loading_info
        return model

    @classmethod
    def _load_pretrained_model_2d(
        cls,
        model,
        state_dict,
        resolved_archive_file,
        pretrained_model_name_or_path,
        ignore_mismatched_sizes=False,
    ):
        # Retrieve missing & unexpected_keys
        model_state_dict = model.state_dict()
        loaded_keys = list(state_dict.keys())

        expected_keys = list(model_state_dict.keys())

        original_loaded_keys = loaded_keys

        missing_keys = list(set(expected_keys) - set(loaded_keys))
        unexpected_keys = list(set(loaded_keys) - set(expected_keys))

        # Make sure we are able to load base models as well as derived models (with heads)
        model_to_load = model

        def _find_mismatched_keys(
            state_dict,
            model_state_dict,
            loaded_keys,
            ignore_mismatched_sizes,
        ):
            mismatched_keys = []
            if ignore_mismatched_sizes:
                for checkpoint_key in loaded_keys:
                    model_key = checkpoint_key

                    if (
                        model_key in model_state_dict
                        and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape
                    ):
                        mismatched_keys.append(
                            (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
                        )
                        del state_dict[checkpoint_key]
            return mismatched_keys

        error_msgs = []
        if state_dict is not None:
            # Whole checkpoint
            mismatched_keys = _find_mismatched_keys(
                state_dict,
                model_state_dict,
                original_loaded_keys,
                ignore_mismatched_sizes,
            )
            # Use PyTorch's load_state_dict with strict=False to handle mismatched keys
            incompatible_keys = model_to_load.load_state_dict(state_dict, strict=False)
            if incompatible_keys.missing_keys:
                error_msgs.append(f"Missing keys: {incompatible_keys.missing_keys}")
            # Note: unexpected_keys are already tracked separately, so we don't include them here

        if len(error_msgs) > 0:
            error_msg = "\n\t".join(error_msgs)
            if "size mismatch" in error_msg:
                error_msg += (
                    "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
                )
            raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")

        if len(unexpected_keys) > 0:
            logger.warning(
                f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when"
                f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are"
                f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task"
                " or with another architecture (e.g. initializing a BertForSequenceClassification model from a"
                " BertForPreTraining model).\n- This IS NOT expected if you are initializing"
                f" {model.__class__.__name__} from the checkpoint of a model that you expect to be exactly"
                " identical (initializing a BertForSequenceClassification model from a"
                " BertForSequenceClassification model)."
            )
        else:
            logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
        if len(missing_keys) > 0:
            logger.warning(
                f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
                f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"
                " TRAIN this model on a down-stream task to be able to use it for predictions and inference."
            )
        elif len(mismatched_keys) == 0:
            logger.info(
                f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"
                f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the"
                f" checkpoint was trained on, you can already use {model.__class__.__name__} for predictions"
                " without further training."
            )
        if len(mismatched_keys) > 0:
            mismatched_warning = "\n".join(
                [
                    f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated"
                    for key, shape1, shape2 in mismatched_keys
                ]
            )
            logger.warning(
                f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
                f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not"
                f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be"
                " able to use it for predictions and inference."
            )

        return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs