showone/models/unet_3d_blocks.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 typing import Any, Dict, Optional, Tuple

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

from diffusers.utils import is_torch_version, logging
from diffusers.models.attention import AdaGroupNorm
from diffusers.models.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
from diffusers.models.resnet import Downsample2D, ResnetBlock2D, TemporalConvLayer, Upsample2D
from diffusers.models.transformer_2d import Transformer2DModel
from diffusers.models.transformer_temporal import TransformerTemporalModel

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


def get_down_block(
    down_block_type,
    num_layers,
    in_channels,
    out_channels,
    temb_channels,
    add_downsample,
    resnet_eps,
    resnet_act_fn,
    transformer_layers_per_block=1,
    num_attention_heads=None,
    resnet_groups=None,
    cross_attention_dim=None,
    downsample_padding=None,
    dual_cross_attention=False,
    use_linear_projection=False,
    only_cross_attention=False,
    upcast_attention=False,
    resnet_time_scale_shift="default",
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
    cross_attention_norm=None,
    attention_head_dim=None,
    downsample_type=None,
):
    # If attn head dim is not defined, we default it to the number of heads
    if attention_head_dim is None:
        logger.warn(
            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
        )
        attention_head_dim = num_attention_heads

    if down_block_type == "DownBlock3D":
        return DownBlock3D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "CrossAttnDownBlock3D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock3D")
        return CrossAttnDownBlock3D(
            num_layers=num_layers,
            transformer_layers_per_block=transformer_layers_per_block,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            cross_attention_dim=cross_attention_dim,
            num_attention_heads=num_attention_heads,
            dual_cross_attention=dual_cross_attention,
            use_linear_projection=use_linear_projection,
            only_cross_attention=only_cross_attention,
            upcast_attention=upcast_attention,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "SimpleCrossAttnDownBlock3D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock3D")
        return SimpleCrossAttnDownBlock3D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
            attention_head_dim=attention_head_dim,
            resnet_time_scale_shift=resnet_time_scale_shift,
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
            only_cross_attention=only_cross_attention,
            cross_attention_norm=cross_attention_norm,
        )
    elif down_block_type == "ResnetDownsampleBlock3D":
        return ResnetDownsampleBlock3D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
        )
    raise ValueError(f"{down_block_type} does not exist.")


def get_up_block(
    up_block_type,
    num_layers,
    in_channels,
    out_channels,
    prev_output_channel,
    temb_channels,
    add_upsample,
    resnet_eps,
    resnet_act_fn,
    transformer_layers_per_block=1,
    num_attention_heads=None,
    resnet_groups=None,
    cross_attention_dim=None,
    dual_cross_attention=False,
    use_linear_projection=False,
    only_cross_attention=False,
    upcast_attention=False,
    resnet_time_scale_shift="default",
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
    cross_attention_norm=None,
    attention_head_dim=None,
    upsample_type=None,
):
    # If attn head dim is not defined, we default it to the number of heads
    if attention_head_dim is None:
        logger.warn(
            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
        )
        attention_head_dim = num_attention_heads

    if up_block_type == "UpBlock3D":
        return UpBlock3D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "CrossAttnUpBlock3D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock3D")
        return CrossAttnUpBlock3D(
            num_layers=num_layers,
            transformer_layers_per_block=transformer_layers_per_block,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
            num_attention_heads=num_attention_heads,
            dual_cross_attention=dual_cross_attention,
            use_linear_projection=use_linear_projection,
            only_cross_attention=only_cross_attention,
            upcast_attention=upcast_attention,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "SimpleCrossAttnUpBlock3D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock3D")
        return SimpleCrossAttnUpBlock3D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
            attention_head_dim=attention_head_dim,
            resnet_time_scale_shift=resnet_time_scale_shift,
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
            only_cross_attention=only_cross_attention,
            cross_attention_norm=cross_attention_norm,
        )
    elif up_block_type == "ResnetUpsampleBlock3D":
        return ResnetUpsampleBlock3D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
        )
    raise ValueError(f"{up_block_type} does not exist.")


class UNetMidBlock3DCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        transformer_layers_per_block: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        num_attention_heads=1,
        output_scale_factor=1.0,
        cross_attention_dim=1280,
        dual_cross_attention=False,
        use_linear_projection=False,
        upcast_attention=False,
    ):
        super().__init__()

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

        # there is always at least one resnet
        resnets = [
            ResnetBlock2D(
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=resnet_groups,
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
            )
        ]
        temp_convs = [
            TemporalConvLayer(
                in_channels,
                in_channels,
                dropout=0.1,
            )
        ]
        attentions = []
        temp_attentions = []

        for _ in range(num_layers):
            attentions.append(
                Transformer2DModel(
                    num_attention_heads,
                    in_channels // num_attention_heads,
                    in_channels=in_channels,
                    num_layers=transformer_layers_per_block,
                    cross_attention_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    use_linear_projection=use_linear_projection,
                    upcast_attention=upcast_attention,
                )
            )
            temp_attentions.append(
                TransformerTemporalModel(
                    num_attention_heads,
                    in_channels // num_attention_heads,
                    in_channels=in_channels,
                    num_layers=1, #todo: transformer_layers_per_block?
                    cross_attention_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                )
            )
            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            temp_convs.append(
                TemporalConvLayer(
                    in_channels,
                    in_channels,
                    dropout=0.1,
                )
            )

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

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

        return hidden_states


class UNetMidBlock3DSimpleCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        attention_head_dim=1,
        output_scale_factor=1.0,
        cross_attention_dim=1280,
        skip_time_act=False,
        only_cross_attention=False,
        cross_attention_norm=None,
    ):
        super().__init__()

        self.has_cross_attention = True

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

        self.num_heads = in_channels // self.attention_head_dim

        # there is always at least one resnet
        resnets = [
            ResnetBlock2D(
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=resnet_groups,
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
                skip_time_act=skip_time_act,
            )
        ]
        temp_convs = [
            TemporalConvLayer(
                in_channels,
                in_channels,
                dropout=0.1,
            )
        ]
        attentions = []
        temp_attentions = []

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

            attentions.append(
                Attention(
                    query_dim=in_channels,
                    cross_attention_dim=in_channels,
                    heads=self.num_heads,
                    dim_head=self.attention_head_dim,
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
                    only_cross_attention=only_cross_attention,
                    cross_attention_norm=cross_attention_norm,
                    processor=processor,
                )
            )
            temp_attentions.append(
                TransformerTemporalModel(
                    self.attention_head_dim,
                    in_channels // self.attention_head_dim,
                    in_channels=in_channels,
                    num_layers=1,
                    cross_attention_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                )
            )
            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                    skip_time_act=skip_time_act,
                )
            )
            temp_convs.append(
                TemporalConvLayer(
                    in_channels,
                    in_channels,
                    dropout=0.1,
                )
            )

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

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        num_frames: int = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}

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

        hidden_states = self.resnets[0](hidden_states, temb)
        hidden_states = self.temp_convs[0](hidden_states, num_frames=num_frames)
        for attn, temp_attn, resnet, temp_conv in zip(
            self.attentions, self.temp_attentions, self.resnets[1:], self.temp_convs[1:]
        ):
            hidden_states = attn(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=mask,
                **cross_attention_kwargs,
            )
            hidden_states = temp_attn(
                hidden_states, num_frames=num_frames, cross_attention_kwargs=cross_attention_kwargs
            ).sample
            hidden_states = resnet(hidden_states, temb)
            hidden_states = temp_conv(hidden_states, num_frames=num_frames)

        return hidden_states


class CrossAttnDownBlock3D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        transformer_layers_per_block: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        num_attention_heads=1,
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        downsample_padding=1,
        add_downsample=True,
        dual_cross_attention=False,
        use_linear_projection=False,
        only_cross_attention=False,
        upcast_attention=False,
    ):
        super().__init__()
        resnets = []
        attentions = []
        temp_attentions = []
        temp_convs = []

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            temp_convs.append(
                TemporalConvLayer(
                    out_channels,
                    out_channels,
                    dropout=0.1,
                )
            )
            attentions.append(
                Transformer2DModel(
                    num_attention_heads,
                    out_channels // num_attention_heads,
                    in_channels=out_channels,
                    num_layers=transformer_layers_per_block,
                    cross_attention_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    use_linear_projection=use_linear_projection,
                    only_cross_attention=only_cross_attention,
                    upcast_attention=upcast_attention,
                )
            )
            temp_attentions.append(
                TransformerTemporalModel(
                    num_attention_heads,
                    out_channels // num_attention_heads,
                    in_channels=out_channels,
                    num_layers=1,
                    cross_attention_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                )
            )
        self.resnets = nn.ModuleList(resnets)
        self.temp_convs = nn.ModuleList(temp_convs)
        self.attentions = nn.ModuleList(attentions)
        self.temp_attentions = nn.ModuleList(temp_attentions)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        num_frames: int = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
    ):
        output_states = ()

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

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

                    return custom_forward

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

            output_states = output_states + (hidden_states,)

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

            output_states = output_states + (hidden_states,)

        return hidden_states, output_states


class DownBlock3D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []
        temp_convs = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            temp_convs.append(
                TemporalConvLayer(
                    out_channels,
                    out_channels,
                    dropout=0.1,
                )
            )

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

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    def forward(self, hidden_states, temb=None, num_frames=1):
        output_states = ()

        for resnet, temp_conv in zip(self.resnets, self.temp_convs):
            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb, use_reentrant=False)
                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(temp_conv), hidden_states, num_frames, use_reentrant=False)
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = temp_conv(hidden_states, num_frames=num_frames)

            output_states = output_states + (hidden_states,)

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

            output_states = output_states + (hidden_states,)

        return hidden_states, output_states


class ResnetDownsampleBlock3D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_downsample=True,
        skip_time_act=False,
    ):
        super().__init__()
        resnets = []
        temp_convs = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                    skip_time_act=skip_time_act,
                )
            )
            temp_convs.append(
                TemporalConvLayer(
                    out_channels,
                    out_channels,
                    dropout=0.1,
                )
            )

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

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_channels,
                        out_channels=out_channels,
                        temb_channels=temb_channels,
                        eps=resnet_eps,
                        groups=resnet_groups,
                        dropout=dropout,
                        time_embedding_norm=resnet_time_scale_shift,
                        non_linearity=resnet_act_fn,
                        output_scale_factor=output_scale_factor,
                        pre_norm=resnet_pre_norm,
                        skip_time_act=skip_time_act,
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    def forward(self, hidden_states, temb=None, num_frames=1):
        output_states = ()

        for resnet, temp_conv in zip(self.resnets, self.temp_convs):
            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb, use_reentrant=False)
                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(temp_conv), hidden_states, num_frames, use_reentrant=False)
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = temp_conv(hidden_states, num_frames=num_frames)

            output_states = output_states + (hidden_states,)

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

            output_states = output_states + (hidden_states,)

        return hidden_states, output_states


class SimpleCrossAttnDownBlock3D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        attention_head_dim=1,
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_downsample=True,
        skip_time_act=False,
        only_cross_attention=False,
        cross_attention_norm=None,
    ):
        super().__init__()

        self.has_cross_attention = True

        resnets = []
        attentions = []
        temp_attentions = []
        temp_convs = []

        self.attention_head_dim = attention_head_dim
        self.num_heads = out_channels // self.attention_head_dim

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                    skip_time_act=skip_time_act,
                )
            )
            temp_convs.append(
                TemporalConvLayer(
                    out_channels,
                    out_channels,
                    dropout=0.1,
                )
            )
            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

            attentions.append(
                Attention(
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
                    dim_head=attention_head_dim,
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
                    only_cross_attention=only_cross_attention,
                    cross_attention_norm=cross_attention_norm,
                    processor=processor,
                )
            )
            temp_attentions.append(
                TransformerTemporalModel(
                    attention_head_dim,
                    out_channels // attention_head_dim,
                    in_channels=out_channels,
                    num_layers=1,
                    cross_attention_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                )
            )
        self.resnets = nn.ModuleList(resnets)
        self.temp_convs = nn.ModuleList(temp_convs)
        self.attentions = nn.ModuleList(attentions)
        self.temp_attentions = nn.ModuleList(temp_attentions)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_channels,
                        out_channels=out_channels,
                        temb_channels=temb_channels,
                        eps=resnet_eps,
                        groups=resnet_groups,
                        dropout=dropout,
                        time_embedding_norm=resnet_time_scale_shift,
                        non_linearity=resnet_act_fn,
                        output_scale_factor=output_scale_factor,
                        pre_norm=resnet_pre_norm,
                        skip_time_act=skip_time_act,
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        num_frames: int = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
    ):
        output_states = ()
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}

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

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

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

                    return custom_forward

                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(temp_conv), hidden_states, num_frames)
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    mask,
                    cross_attention_kwargs,
                )[0]
                hidden_states = temp_attn(
                    hidden_states, num_frames=num_frames, cross_attention_kwargs=cross_attention_kwargs
                ).sample
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = temp_conv(hidden_states, num_frames=num_frames)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=mask,
                    **cross_attention_kwargs,
                )
                hidden_states = temp_attn(
                    hidden_states, num_frames=num_frames, cross_attention_kwargs=cross_attention_kwargs
                ).sample

            output_states = output_states + (hidden_states,)

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

            output_states = output_states + (hidden_states,)

        return hidden_states, output_states


class CrossAttnUpBlock3D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        transformer_layers_per_block: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        num_attention_heads=1,
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
        dual_cross_attention=False,
        use_linear_projection=False,
        only_cross_attention=False,
        upcast_attention=False,
    ):
        super().__init__()
        resnets = []
        temp_convs = []
        attentions = []
        temp_attentions = []

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads

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

            resnets.append(
                ResnetBlock2D(
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            temp_convs.append(
                TemporalConvLayer(
                    out_channels,
                    out_channels,
                    dropout=0.1,
                )
            )
            attentions.append(
                Transformer2DModel(
                    num_attention_heads,
                    out_channels // num_attention_heads,
                    in_channels=out_channels,
                    num_layers=transformer_layers_per_block,
                    cross_attention_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    use_linear_projection=use_linear_projection,
                    only_cross_attention=only_cross_attention,
                    upcast_attention=upcast_attention,
                )
            )
            temp_attentions.append(
                TransformerTemporalModel(
                    num_attention_heads,
                    out_channels // num_attention_heads,
                    in_channels=out_channels,
                    num_layers=1,
                    cross_attention_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                )
            )
        self.resnets = nn.ModuleList(resnets)
        self.temp_convs = nn.ModuleList(temp_convs)
        self.attentions = nn.ModuleList(attentions)
        self.temp_attentions = nn.ModuleList(temp_attentions)

        if add_upsample:
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

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

            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

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

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

        return hidden_states


class UpBlock3D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []
        temp_convs = []

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

            resnets.append(
                ResnetBlock2D(
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            temp_convs.append(
                TemporalConvLayer(
                    out_channels,
                    out_channels,
                    dropout=0.1,
                )
            )

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

        if add_upsample:
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

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

            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb, use_reentrant=False)
                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(temp_conv), hidden_states, num_frames, use_reentrant=False)
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = temp_conv(hidden_states, num_frames=num_frames)

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

        return hidden_states


class ResnetUpsampleBlock3D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_upsample=True,
        skip_time_act=False,
    ):
        super().__init__()
        resnets = []
        temp_convs = []

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

            resnets.append(
                ResnetBlock2D(
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                    skip_time_act=skip_time_act,
                )
            )
            temp_convs.append(
                TemporalConvLayer(
                    out_channels,
                    out_channels,
                    dropout=0.1,
                )
            )

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

        if add_upsample:
            self.upsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_channels,
                        out_channels=out_channels,
                        temb_channels=temb_channels,
                        eps=resnet_eps,
                        groups=resnet_groups,
                        dropout=dropout,
                        time_embedding_norm=resnet_time_scale_shift,
                        non_linearity=resnet_act_fn,
                        output_scale_factor=output_scale_factor,
                        pre_norm=resnet_pre_norm,
                        skip_time_act=skip_time_act,
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

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

            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb, use_reentrant=False)
                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(temp_conv), hidden_states, num_frames, use_reentrant=False)
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = temp_conv(hidden_states, num_frames=num_frames)

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

        return hidden_states


class SimpleCrossAttnUpBlock3D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        attention_head_dim=1,
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
        skip_time_act=False,
        only_cross_attention=False,
        cross_attention_norm=None,
    ):
        super().__init__()
        resnets = []
        temp_convs = []
        attentions = []
        temp_attentions = []

        self.has_cross_attention = True
        self.attention_head_dim = attention_head_dim

        self.num_heads = out_channels // self.attention_head_dim

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

            resnets.append(
                ResnetBlock2D(
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                    skip_time_act=skip_time_act,
                )
            )
            temp_convs.append(
                TemporalConvLayer(
                    out_channels,
                    out_channels,
                    dropout=0.1,
                )
            )

            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

            attentions.append(
                Attention(
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
                    dim_head=self.attention_head_dim,
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
                    only_cross_attention=only_cross_attention,
                    cross_attention_norm=cross_attention_norm,
                    processor=processor,
                )
            )
            temp_attentions.append(
                TransformerTemporalModel(
                    attention_head_dim,
                    out_channels // attention_head_dim,
                    in_channels=out_channels,
                    num_layers=1,
                    cross_attention_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                )
            )
        self.resnets = nn.ModuleList(resnets)
        self.temp_convs = nn.ModuleList(temp_convs)
        self.attentions = nn.ModuleList(attentions)
        self.temp_attentions = nn.ModuleList(temp_attentions)

        if add_upsample:
            self.upsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_channels,
                        out_channels=out_channels,
                        temb_channels=temb_channels,
                        eps=resnet_eps,
                        groups=resnet_groups,
                        dropout=dropout,
                        time_embedding_norm=resnet_time_scale_shift,
                        non_linearity=resnet_act_fn,
                        output_scale_factor=output_scale_factor,
                        pre_norm=resnet_pre_norm,
                        skip_time_act=skip_time_act,
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        upsample_size: Optional[int] = None,
        num_frames: int = 1,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}

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

        for resnet, temp_conv, attn, temp_attn in zip(
            self.resnets, self.temp_convs, self.attentions, self.temp_attentions
        ):
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(temp_conv), hidden_states, num_frames)
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    mask,
                    cross_attention_kwargs,
                )[0]
                hidden_states = temp_attn(
                    hidden_states, num_frames=num_frames, cross_attention_kwargs=cross_attention_kwargs
                ).sample
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = temp_conv(hidden_states, num_frames=num_frames)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=mask,
                    **cross_attention_kwargs,
                )
                hidden_states = temp_attn(
                    hidden_states, num_frames=num_frames, cross_attention_kwargs=cross_attention_kwargs
                ).sample

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

        return hidden_states