consisti2v/models/videoldm_unet.py

import os
import re
from typing import Optional, Tuple, Union, Dict, List, Any
from einops import rearrange, repeat

import torch
import torch.nn as nn
from diffusers.loaders import UNet2DConditionLoadersMixin
from diffusers.models import ModelMixin
from diffusers.models.unet_2d_condition import UNet2DConditionOutput
from diffusers.models.unet_2d_blocks import UNetMidBlock2DCrossAttn, UNetMidBlock2DSimpleCrossAttn
from diffusers.models.embeddings import (
    GaussianFourierProjection,
    ImageHintTimeEmbedding,
    ImageProjection,
    ImageTimeEmbedding,
    PositionNet,
    TextImageProjection,
    TextImageTimeEmbedding,
    TextTimeEmbedding,
    TimestepEmbedding,
    Timesteps,
)
from diffusers.models.attention_processor import (
    ADDED_KV_ATTENTION_PROCESSORS,
    CROSS_ATTENTION_PROCESSORS,
    AttentionProcessor,
    AttnAddedKVProcessor,
    AttnProcessor,
)
from diffusers.models.activations import get_activation
from diffusers.configuration_utils import register_to_config, ConfigMixin
from diffusers.models.modeling_utils import load_state_dict, load_model_dict_into_meta
from diffusers.utils import (
    CONFIG_NAME,
    DIFFUSERS_CACHE,
    FLAX_WEIGHTS_NAME,
    HF_HUB_OFFLINE,
    SAFETENSORS_WEIGHTS_NAME,
    WEIGHTS_NAME,
    _add_variant,
    _get_model_file,
    deprecate,
    is_accelerate_available,
    is_torch_version,
    logging,
)
from diffusers import __version__

if is_torch_version(">=", "1.9.0"):
    _LOW_CPU_MEM_USAGE_DEFAULT = True
else:
    _LOW_CPU_MEM_USAGE_DEFAULT = False


if is_accelerate_available():
    import accelerate
    from accelerate.utils import set_module_tensor_to_device
    from accelerate.utils.versions import is_torch_version



from .videoldm_unet_blocks import get_down_block, get_up_block, VideoLDMUNetMidBlock2DCrossAttn

logger = logging.get_logger(__name__)


class VideoLDMUNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
    _supports_gradient_checkpointing = True
    @register_to_config
    def __init__(
        self,
        sample_size: Optional[int] = None,
        in_channels: int = 4,
        out_channels: int = 4,
        center_input_sample: bool = False,
        flip_sin_to_cos: bool = True,
        freq_shift: int = 0,
        down_block_types: Tuple[str] = (
            "CrossAttnDownBlock2D", # -> VideoLDMDownBlock
            "CrossAttnDownBlock2D", # -> VideoLDMDownBlock
            "CrossAttnDownBlock2D", # -> VideoLDMDownBlock
            "DownBlock2D",
        ),
        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
        up_block_types: Tuple[str] = (
            "UpBlock2D",
            "CrossAttnUpBlock2D", # -> VideoLDMUpBlock
            "CrossAttnUpBlock2D", # -> VideoLDMUpBlock
            "CrossAttnUpBlock2D", # -> VideoLDMUpBlock
        ),
        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,
        dropout: float = 0.0,
        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,
        attention_type: str = "default",
        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,
        # additional
        use_temporal: bool = True,
        n_frames: int = 8,
        n_temp_heads: int = 8,
        first_frame_condition_mode: str = "none",
        augment_temporal_attention: bool = False,
        temp_pos_embedding: str = "sinusoidal",
        use_frame_stride_condition: bool = False,
    ):
        super().__init__()

        rotary_emb = False
        if temp_pos_embedding == "rotary":
            # from rotary_embedding_torch import RotaryEmbedding
            # rotary_emb = RotaryEmbedding(32)
            # self.rotary_emb = rotary_emb
            rotary_emb = True
            self.rotary_emb = rotary_emb

        self.use_temporal = use_temporal
        self.augment_temporal_attention = augment_temporal_attention

        assert first_frame_condition_mode in ["none", "concat", "conv2d", "input_only"], f"first_frame_condition_mode: {first_frame_condition_mode} must be one of ['none', 'concat', 'conv2d', 'input_only']"
        self.first_frame_condition_mode = first_frame_condition_mode
        latent_channels = in_channels

        self.sample_size = sample_size

        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."
            )
        
        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,
        )

        self.use_frame_stride_condition = use_frame_stride_condition
        if self.use_frame_stride_condition:
            self.frame_stride_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,
            )
            # zero init
            nn.init.zeros_(self.frame_stride_embedding.linear_2.weight)
            nn.init.zeros_(self.frame_stride_embedding.linear_2.bias)

        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,
                attention_type=attention_type,
                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,
                dropout=dropout,
                # additional
                use_temporal=use_temporal,
                augment_temporal_attention=augment_temporal_attention,
                n_frames=n_frames,
                n_temp_heads=n_temp_heads,
                first_frame_condition_mode=first_frame_condition_mode,
                latent_channels=latent_channels,
                rotary_emb=rotary_emb,
            )
            self.down_blocks.append(down_block)

        # mid
        if mid_block_type == "UNetMidBlock2DCrossAttn":
            self.mid_block = VideoLDMUNetMidBlock2DCrossAttn(
                transformer_layers_per_block=transformer_layers_per_block[-1],
                in_channels=block_out_channels[-1],
                temb_channels=blocks_time_embed_dim,
                dropout=dropout,
                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,
                attention_type=attention_type,
                # additional
                use_temporal=use_temporal,
                n_frames=n_frames,
                first_frame_condition_mode=first_frame_condition_mode,
                latent_channels=latent_channels,
            )
        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
                in_channels=block_out_channels[-1],
                temb_channels=blocks_time_embed_dim,
                dropout=dropout,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
                cross_attention_dim=cross_attention_dim[-1],
                attention_head_dim=attention_head_dim[-1],
                resnet_groups=norm_num_groups,
                resnet_time_scale_shift=resnet_time_scale_shift,
                skip_time_act=resnet_skip_time_act,
                only_cross_attention=mid_block_only_cross_attention,
                cross_attention_norm=cross_attention_norm,
            )
        elif mid_block_type is None:
            self.mid_block = None
        else:
            raise ValueError(f"unknown mid_block_type : {mid_block_type}")

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

        # up
        reversed_block_out_channels = list(reversed(block_out_channels))
        reversed_num_attention_heads = list(reversed(num_attention_heads))
        reversed_layers_per_block = list(reversed(layers_per_block))
        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
        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,
                attention_type=attention_type,
                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,
                dropout=dropout,
                # additional
                use_temporal=use_temporal,
                augment_temporal_attention=augment_temporal_attention,
                n_frames=n_frames,
                n_temp_heads=n_temp_heads,
                first_frame_condition_mode=first_frame_condition_mode,
                latent_channels=latent_channels,
                rotary_emb=rotary_emb,
            )
            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, "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

    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.
        """
        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)

    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 hasattr(module, "gradient_checkpointing"):
            module.gradient_checkpointing = value

    def forward(
        self,
        sample: torch.FloatTensor,
        timestep: Union[torch.Tensor, float, int],
        encoder_hidden_states: torch.Tensor,
        class_labels: Optional[torch.Tensor] = None,
        timestep_cond: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
        mid_block_additional_residual: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        return_dict: bool = True,
        # additional
        first_frame_latents: Optional[torch.Tensor] = None,
        frame_stride: Optional[Union[torch.Tensor, float, int]] = None,
    ) -> Union[UNet2DConditionOutput, Tuple]:
        # reshape video data
        assert sample.dim() == 5, f"Expected hidden_states to have ndim=5, but got ndim={sample.dim()}."
        video_length = sample.shape[2]
        
        if first_frame_latents is not None:
            assert self.config.first_frame_condition_mode != "none", "first_frame_latents is not None, but first_frame_condition_mode is 'none'."
        
        if self.config.first_frame_condition_mode != "none":
            sample = torch.cat([first_frame_latents, sample], dim=2)
            video_length += 1
        
        # copy conditioning embeddings for cross attention
        if encoder_hidden_states is not None:
            encoder_hidden_states = repeat(encoder_hidden_states, 'b n c -> (b f) n c', f=video_length)
        
        sample = rearrange(sample, "b c f h w -> (b f) c h w")

        # By default samples have to be AT least a multiple of the overall upsampling factor.
        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
        # However, the upsampling interpolation output size can be forced to fit any upsampling size
        # on the fly if necessary.
        default_overall_up_factor = 2**self.num_upsamplers

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

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

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

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

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

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

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

        t_emb = self.time_proj(timesteps)

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

        emb = self.time_embedding(t_emb, timestep_cond)

        if self.use_frame_stride_condition:
            if not torch.is_tensor(frame_stride):
                # 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
                frame_stride = torch.tensor([frame_stride], dtype=dtype, device=sample.device)
            elif len(frame_stride.shape) == 0:
                frame_stride = frame_stride[None].to(sample.device)

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

            fs_emb = self.time_proj(frame_stride)

            # `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.
            fs_emb = fs_emb.to(dtype=sample.dtype)

            fs_emb = self.frame_stride_embedding(fs_emb, timestep_cond)
            emb = emb + fs_emb

        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

        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)

        # 2.5 GLIGEN position net
        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
            cross_attention_kwargs = cross_attention_kwargs.copy()
            gligen_args = cross_attention_kwargs.pop("gligen")
            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}

        # 3. down
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0

        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 downsample_block in 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,
                    first_frame_latents=first_frame_latents,
                    **additional_residuals,
                )
            else:
                sample, res_samples = downsample_block(hidden_states=sample, temb=emb, scale=lora_scale, first_frame_latents=first_frame_latents,)

                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

        # 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,
                # additional
                first_frame_latents=first_frame_latents,
            )
            # To support T2I-Adapter-XL
            if (
                is_adapter
                and len(down_block_additional_residuals) > 0
                and sample.shape == down_block_additional_residuals[0].shape
            ):
                sample += down_block_additional_residuals.pop(0)

        if is_controlnet:
            sample = sample + mid_block_additional_residual

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

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

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

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

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

        sample = rearrange(sample, "(b f) c h w -> b c f h w", f=video_length)
        if self.config.first_frame_condition_mode != "none":
            sample = sample[:, :, 1:, :, :]

        if not return_dict:
            return (sample,)

        return UNet2DConditionOutput(sample=sample)

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):

        kwargs.pop("low_cpu_mem_usage", False)
        kwargs.pop("device_map", None)
        
        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 = None
        max_memory = kwargs.pop("max_memory", None)
        offload_folder = kwargs.pop("offload_folder", None)
        offload_state_dict = kwargs.pop("offload_state_dict", False)
        low_cpu_mem_usage = False
        variant = kwargs.pop("variant", None)
        use_safetensors = kwargs.pop("use_safetensors", None)

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

        if low_cpu_mem_usage and not is_accelerate_available():
            low_cpu_mem_usage = False
            logger.warning(
                "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
                " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
                " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
                " install accelerate\n```\n."
            )

        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`."
            )

        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
            raise NotImplementedError(
                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
                " `low_cpu_mem_usage=False`."
            )

        if low_cpu_mem_usage is False and device_map is not None:
            raise ValueError(
                f"You cannot set `low_cpu_mem_usage` to `False` while using device_map={device_map} for loading and"
                " dispatching. Please make sure to set `low_cpu_mem_usage=True`."
            )

        # 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,
        )

        # load model
        model_file = None
        if from_flax:
            model_file = _get_model_file(
                pretrained_model_name_or_path,
                weights_name=FLAX_WEIGHTS_NAME,
                cache_dir=cache_dir,
                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,
                user_agent=user_agent,
                commit_hash=commit_hash,
            )
            model = cls.from_config(config, **unused_kwargs)

            # Convert the weights
            from diffusers.models.modeling_pytorch_flax_utils import load_flax_checkpoint_in_pytorch_model

            model = load_flax_checkpoint_in_pytorch_model(model, model_file)
        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,
                        resume_download=resume_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,
                    resume_download=resume_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,
                )

            if low_cpu_mem_usage:
                # Instantiate model with empty weights
                with accelerate.init_empty_weights():
                    model = cls.from_config(config, **unused_kwargs)

                # if device_map is None, load the state dict and move the params from meta device to the cpu
                if device_map is None:
                    param_device = "cpu"
                    state_dict = load_state_dict(model_file, variant=variant)
                    model._convert_deprecated_attention_blocks(state_dict)
                    # move the params from meta device to cpu
                    missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())
                    if len(missing_keys) > 0:
                        raise ValueError(
                            f"Cannot load {cls} from {pretrained_model_name_or_path} because the following keys are"
                            f" missing: \n {', '.join(missing_keys)}. \n Please make sure to pass"
                            " `low_cpu_mem_usage=False` and `device_map=None` if you want to randomly initialize"
                            " those weights or else make sure your checkpoint file is correct."
                        )

                    unexpected_keys = load_model_dict_into_meta(
                        model,
                        state_dict,
                        device=param_device,
                        dtype=torch_dtype,
                        model_name_or_path=pretrained_model_name_or_path,
                    )

                    if cls._keys_to_ignore_on_load_unexpected is not None:
                        for pat in cls._keys_to_ignore_on_load_unexpected:
                            unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]

                    if len(unexpected_keys) > 0:
                        logger.warn(
                            f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
                        )

                else:  # else let accelerate handle loading and dispatching.
                    # Load weights and dispatch according to the device_map
                    # by default the device_map is None and the weights are loaded on the CPU
                    try:
                        accelerate.load_checkpoint_and_dispatch(
                            model,
                            model_file,
                            device_map,
                            max_memory=max_memory,
                            offload_folder=offload_folder,
                            offload_state_dict=offload_state_dict,
                            dtype=torch_dtype,
                        )
                    except AttributeError as e:
                        # When using accelerate loading, we do not have the ability to load the state
                        # dict and rename the weight names manually. Additionally, accelerate skips
                        # torch loading conventions and directly writes into `module.{_buffers, _parameters}`
                        # (which look like they should be private variables?), so we can't use the standard hooks
                        # to rename parameters on load. We need to mimic the original weight names so the correct
                        # attributes are available. After we have loaded the weights, we convert the deprecated
                        # names to the new non-deprecated names. Then we _greatly encourage_ the user to convert
                        # the weights so we don't have to do this again.

                        if "'Attention' object has no attribute" in str(e):
                            logger.warn(
                                f"Taking `{str(e)}` while using `accelerate.load_checkpoint_and_dispatch` to mean {pretrained_model_name_or_path}"
                                " was saved with deprecated attention block weight names. We will load it with the deprecated attention block"
                                " names and convert them on the fly to the new attention block format. Please re-save the model after this conversion,"
                                " so we don't have to do the on the fly renaming in the future. If the model is from a hub checkpoint,"
                                " please also re-upload it or open a PR on the original repository."
                            )
                            model._temp_convert_self_to_deprecated_attention_blocks()
                            accelerate.load_checkpoint_and_dispatch(
                                model,
                                model_file,
                                device_map,
                                max_memory=max_memory,
                                offload_folder=offload_folder,
                                offload_state_dict=offload_state_dict,
                                dtype=torch_dtype,
                            )
                            model._undo_temp_convert_self_to_deprecated_attention_blocks()
                        else:
                            raise e

                loading_info = {
                    "missing_keys": [],
                    "unexpected_keys": [],
                    "mismatched_keys": [],
                    "error_msgs": [],
                }
            else:
                model = cls.from_config(config, **unused_kwargs)

                state_dict = load_state_dict(model_file, variant=variant)
                model._convert_deprecated_attention_blocks(state_dict)

                model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model(
                    model,
                    state_dict,
                    model_file,
                    pretrained_model_name_or_path,
                    ignore_mismatched_sizes=ignore_mismatched_sizes,
                )

                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)

        m, u = loading_info["missing_keys"], loading_info["unexpected_keys"]
        logger.info(f"### missing keys: {len(m)}; unexpected keys: {len(u)};")
        # print(f"### missing keys:\n{m}\n### unexpected keys:\n{u}\n")

        spatial_params = [p.numel() if "conv3ds" not in n and "tempo_attns" not in n else 0 for n, p in model.named_parameters()]
        tconv_params = [p.numel() if "conv3ds." in n else 0 for n, p in model.named_parameters()]
        tattn_params = [p.numel() if "tempo_attns." in n else 0 for n, p in model.named_parameters()]
        tffconv_params = [p.numel() if "first_frame_conv." in n else 0 for n, p in model.named_parameters()]
        logger.info(f"### First Frame Convolution Layer Parameters: {sum(tffconv_params) / 1e6} M")
        logger.info(f"### Spatial UNet Parameters: {sum(spatial_params) / 1e6} M")
        logger.info(f"### Temporal Convolution Module Parameters: {sum(tconv_params) / 1e6} M")
        logger.info(f"### Temporal Attention Module Parameters: {sum(tattn_params) / 1e6} M")

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

        return model

if __name__ == "__main__":
    # test
    from diffusers import AutoencoderKL, DDIMScheduler
    from transformers import CLIPTextModel, CLIPTokenizer
    from consisti2v.pipelines.pipeline_animation import AnimationPipeline
    from consisti2v.pipelines.pipeline_conditional_animation import ConditionalAnimationPipeline
    from consisti2v.utils.util import save_videos_grid

    pretrained_model_path = "models/StableDiffusion/stable-diffusion-v1-5"
    prompt = "apply eye makeup"
    first_frame_path = "/ML-A100/home/weiming/datasets/UCF/frames/v_ApplyEyeMakeup_g01_c01_frame_90.jpg"

    tokenizer    = CLIPTokenizer.from_pretrained(pretrained_model_path, subfolder="tokenizer", use_safetensors=True)
    text_encoder = CLIPTextModel.from_pretrained(pretrained_model_path, subfolder="text_encoder")
    vae          = AutoencoderKL.from_pretrained(pretrained_model_path, subfolder="vae", use_safetensors=True)            
    unet         = VideoLDMUNet3DConditionModel.from_pretrained(
        pretrained_model_path,
        subfolder="unet",
        use_safetensors=True
    )

    noise_scheduler_kwargs = {
        "num_train_timesteps": 1000,
        "beta_start":          0.00085,
        "beta_end":            0.012,
        "beta_schedule":       "linear",
        "steps_offset":        1,
        "clip_sample":         False,
    }
    noise_scheduler = DDIMScheduler(**noise_scheduler_kwargs)
    # latent = torch.randn(1, 4, 8, 64, 64).to("cuda")
    # text_embedding = torch.randn(1, 77, 768).to("cuda")
    # timestep = torch.randint(0, 1000, (1,)).to("cuda").squeeze(0)
    # output = unet(latent, timestep, text_embedding)

    pipeline = ConditionalAnimationPipeline(
        unet=unet, vae=vae, tokenizer=tokenizer, text_encoder=text_encoder, scheduler=noise_scheduler,
    ).to("cuda")
    sample = pipeline(
        prompt,
        num_inference_steps = 25,
        guidance_scale      = 8.,
        video_length        = 8,
        height              = 256,
        width               =  256,
        first_frame_paths   = first_frame_path,
    ).videos
    print(sample.shape)
    save_videos_grid(sample, f"samples/videoldm.gif")