ppdiffusers/models/unet_2d_blocks.py

# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
# Copyright 2022 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.
import numpy as np
import paddle
from paddle import nn
from paddle.distributed.fleet.utils import recompute

from .attention import AttentionBlock, DualTransformer2DModel, Transformer2DModel
from .cross_attention import CrossAttention, CrossAttnAddedKVProcessor
from .resnet import (
    Downsample2D,
    FirDownsample2D,
    FirUpsample2D,
    ResnetBlock2D,
    Upsample2D,
)


def get_down_block(
    down_block_type,
    num_layers,
    in_channels,
    out_channels,
    temb_channels,
    add_downsample,
    resnet_eps,
    resnet_act_fn,
    attn_num_head_channels,
    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",
):
    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
    if down_block_type == "DownBlock2D":
        return DownBlock2D(
            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 == "ResnetDownsampleBlock2D":
        return ResnetDownsampleBlock2D(
            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,
        )
    elif down_block_type == "AttnDownBlock2D":
        return AttnDownBlock2D(
            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,
            attn_num_head_channels=attn_num_head_channels,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "CrossAttnDownBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
        return CrossAttnDownBlock2D(
            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,
            cross_attention_dim=cross_attention_dim,
            attn_num_head_channels=attn_num_head_channels,
            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 == "SimpleCrossAttnDownBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D")
        return SimpleCrossAttnDownBlock2D(
            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,
            attn_num_head_channels=attn_num_head_channels,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "SkipDownBlock2D":
        return SkipDownBlock2D(
            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,
            downsample_padding=downsample_padding,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "AttnSkipDownBlock2D":
        return AttnSkipDownBlock2D(
            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,
            downsample_padding=downsample_padding,
            attn_num_head_channels=attn_num_head_channels,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "DownEncoderBlock2D":
        return DownEncoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_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 == "AttnDownEncoderBlock2D":
        return AttnDownEncoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            attn_num_head_channels=attn_num_head_channels,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    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,
    attn_num_head_channels,
    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",
):
    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
    if up_block_type == "UpBlock2D":
        return UpBlock2D(
            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 == "ResnetUpsampleBlock2D":
        return ResnetUpsampleBlock2D(
            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 == "CrossAttnUpBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
        return CrossAttnUpBlock2D(
            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,
            attn_num_head_channels=attn_num_head_channels,
            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 == "SimpleCrossAttnUpBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D")
        return SimpleCrossAttnUpBlock2D(
            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,
            attn_num_head_channels=attn_num_head_channels,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "AttnUpBlock2D":
        return AttnUpBlock2D(
            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,
            attn_num_head_channels=attn_num_head_channels,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "SkipUpBlock2D":
        return SkipUpBlock2D(
            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_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "AttnSkipUpBlock2D":
        return AttnSkipUpBlock2D(
            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,
            attn_num_head_channels=attn_num_head_channels,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "UpDecoderBlock2D":
        return UpDecoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_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 == "AttnUpDecoderBlock2D":
        return AttnUpDecoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            attn_num_head_channels=attn_num_head_channels,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    raise ValueError(f"{up_block_type} does not exist.")


class UNetMidBlock2D(nn.Layer):
    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,
        add_attention: bool = True,
        attn_num_head_channels=1,
        output_scale_factor=1.0,
    ):
        super().__init__()

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

        # 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,
            )
        ]
        attentions = []

        for _ in range(num_layers):
            if self.add_attention:
                attentions.append(
                    AttentionBlock(
                        in_channels,
                        num_head_channels=attn_num_head_channels,
                        rescale_output_factor=output_scale_factor,
                        eps=resnet_eps,
                        norm_num_groups=resnet_groups,
                    )
                )
            else:
                attentions.append(None)

            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,
                )
            )

        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

    def forward(self, hidden_states, temb=None):
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            if attn is not None:
                hidden_states = attn(hidden_states)
            hidden_states = resnet(hidden_states, temb)

        return hidden_states


class UNetMidBlock2DCrossAttn(nn.Layer):
    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,
        attn_num_head_channels=1,
        output_scale_factor=1.0,
        cross_attention_dim=1280,
        dual_cross_attention=False,
        use_linear_projection=False,
        only_cross_attention=False,
        upcast_attention=False,
    ):
        super().__init__()

        self.has_cross_attention = True
        self.attn_num_head_channels = attn_num_head_channels
        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,
            )
        ]
        attentions = []

        for _ in range(num_layers):
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
                        attn_num_head_channels,
                        in_channels // attn_num_head_channels,
                        in_channels=in_channels,
                        num_layers=1,
                        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,
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
                        attn_num_head_channels,
                        in_channels // attn_num_head_channels,
                        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,
                )
            )

        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

    def forward(
        self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None, cross_attention_kwargs=None
    ):
        # TODO(Patrick, William) - attention_mask is currently not used. Implement once used
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            hidden_states = attn(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                cross_attention_kwargs=cross_attention_kwargs,
            ).sample
            hidden_states = resnet(hidden_states, temb)

        return hidden_states


class UNetMidBlock2DSimpleCrossAttn(nn.Layer):
    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,
        attn_num_head_channels=1,
        output_scale_factor=1.0,
        cross_attention_dim=1280,
    ):
        super().__init__()

        self.has_cross_attention = True

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

        self.num_heads = in_channels // self.attn_num_head_channels

        # 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,
            )
        ]
        attentions = []

        for _ in range(num_layers):
            attentions.append(
                CrossAttention(
                    query_dim=in_channels,
                    cross_attention_dim=in_channels,
                    heads=self.num_heads,
                    dim_head=attn_num_head_channels,
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
                    processor=CrossAttnAddedKVProcessor(),
                )
            )
            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,
                )
            )

        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

    def set_attention_slice(self, slice_size):
        head_dims = self.attn_num_head_channels
        head_dims = [head_dims] if isinstance(head_dims, int) else head_dims
        if slice_size is not None and any(dim % slice_size != 0 for dim in head_dims):
            raise ValueError(
                f"Make sure slice_size {slice_size} is a common divisor of "
                f"the number of heads used in cross_attention: {head_dims}"
            )
        if slice_size is not None and slice_size > min(head_dims):
            raise ValueError(
                f"slice_size {slice_size} has to be smaller or equal to "
                f"the lowest number of heads used in cross_attention: min({head_dims}) = {min(head_dims)}"
            )

        for attn in self.attentions:
            attn._set_attention_slice(slice_size)

    def forward(
        self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None, cross_attention_kwargs=None
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            # attn
            hidden_states = attn(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=attention_mask,
                **cross_attention_kwargs,
            )
            # resnet
            hidden_states = resnet(hidden_states, temb)

        return hidden_states


class AttnDownBlock2D(nn.Layer):
    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,
        attn_num_head_channels=1,
        attention_type="default",
        output_scale_factor=1.0,
        downsample_padding=1,
        add_downsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.attention_type = attention_type

        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,
                )
            )
            attentions.append(
                AttentionBlock(
                    out_channels,
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
                    norm_num_groups=resnet_groups,
                )
            )

        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

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

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

        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb)
            hidden_states = attn(hidden_states)
            output_states += (hidden_states,)

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

            output_states += (hidden_states,)

        return hidden_states, output_states


class CrossAttnDownBlock2D(nn.Layer):
    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,
        attn_num_head_channels=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 = []

        self.has_cross_attention = True
        self.attn_num_head_channels = attn_num_head_channels

        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,
                )
            )
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
                        attn_num_head_channels,
                        out_channels // attn_num_head_channels,
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                        use_linear_projection=use_linear_projection,
                        upcast_attention=upcast_attention,
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
                        attn_num_head_channels,
                        out_channels // attn_num_head_channels,
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
                )
        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

        if add_downsample:
            self.downsamplers = nn.LayerList(
                [
                    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, encoder_hidden_states=None, attention_mask=None, cross_attention_kwargs=None
    ):
        # TODO(Patrick, William) - attention mask is not used
        output_states = ()

        for resnet, attn in zip(self.resnets, self.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)[0]  # move [0]
                        else:
                            return module(*inputs)

                    return custom_forward

                hidden_states = recompute(create_custom_forward(resnet), hidden_states, temb)
                hidden_states = recompute(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    cross_attention_kwargs,
                )  # [0]
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                ).sample
            output_states += (hidden_states,)

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

            output_states += (hidden_states,)

        return hidden_states, output_states


class DownBlock2D(nn.Layer):
    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 = []

        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,
                )
            )

        self.resnets = nn.LayerList(resnets)

        if add_downsample:
            self.downsamplers = nn.LayerList(
                [
                    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):
        output_states = ()

        for resnet in self.resnets:
            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

                hidden_states = recompute(create_custom_forward(resnet), hidden_states, temb)
            else:
                hidden_states = resnet(hidden_states, temb)

            output_states += (hidden_states,)

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

            output_states += (hidden_states,)

        return hidden_states, output_states


class DownEncoderBlock2D(nn.Layer):
    def __init__(
        self,
        in_channels: int,
        out_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 = []

        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=None,
                    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,
                )
            )

        self.resnets = nn.LayerList(resnets)

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

    def forward(self, hidden_states):
        for resnet in self.resnets:
            hidden_states = resnet(hidden_states, temb=None)

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

        return hidden_states


class AttnDownEncoderBlock2D(nn.Layer):
    def __init__(
        self,
        in_channels: int,
        out_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,
        attn_num_head_channels=1,
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []
        attentions = []

        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=None,
                    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,
                )
            )
            attentions.append(
                AttentionBlock(
                    out_channels,
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
                    norm_num_groups=resnet_groups,
                )
            )

        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

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

    def forward(self, hidden_states):
        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb=None)
            hidden_states = attn(hidden_states)

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

        return hidden_states


class AttnSkipDownBlock2D(nn.Layer):
    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_pre_norm: bool = True,
        attn_num_head_channels=1,
        output_scale_factor=np.sqrt(2.0),
        downsample_padding=1,
        add_downsample=True,
    ):
        super().__init__()
        self.attentions = nn.LayerList([])
        self.resnets = nn.LayerList([])

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            self.resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min(in_channels // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    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,
                )
            )
            self.attentions.append(
                AttentionBlock(
                    out_channels,
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
                )
            )

        if add_downsample:
            self.resnet_down = ResnetBlock2D(
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                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,
                use_in_shortcut=True,
                down=True,
                kernel="fir",
            )
            self.downsamplers = nn.LayerList([FirDownsample2D(out_channels, out_channels=out_channels)])
            self.skip_conv = nn.Conv2D(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
        else:
            self.resnet_down = None
            self.downsamplers = None
            self.skip_conv = None

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

        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb)
            hidden_states = attn(hidden_states)
            output_states += (hidden_states,)

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

            hidden_states = self.skip_conv(skip_sample) + hidden_states

            output_states += (hidden_states,)

        return hidden_states, output_states, skip_sample


class SkipDownBlock2D(nn.Layer):
    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_pre_norm: bool = True,
        output_scale_factor=np.sqrt(2.0),
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        self.resnets = nn.LayerList([])

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            self.resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min(in_channels // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    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,
                )
            )

        if add_downsample:
            self.resnet_down = ResnetBlock2D(
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                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,
                use_in_shortcut=True,
                down=True,
                kernel="fir",
            )
            self.downsamplers = nn.LayerList([FirDownsample2D(out_channels, out_channels=out_channels)])
            self.skip_conv = nn.Conv2D(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
        else:
            self.resnet_down = None
            self.downsamplers = None
            self.skip_conv = None

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

        for resnet in self.resnets:
            hidden_states = resnet(hidden_states, temb)
            output_states += (hidden_states,)

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

            hidden_states = self.skip_conv(skip_sample) + hidden_states

            output_states += (hidden_states,)

        return hidden_states, output_states, skip_sample


class ResnetDownsampleBlock2D(nn.Layer):
    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,
    ):
        super().__init__()
        resnets = []

        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,
                )
            )

        self.resnets = nn.LayerList(resnets)

        if add_downsample:
            self.downsamplers = nn.LayerList(
                [
                    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,
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

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

        for resnet in self.resnets:
            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

                hidden_states = recompute(create_custom_forward(resnet), hidden_states, temb)
            else:
                hidden_states = resnet(hidden_states, temb)

            output_states += (hidden_states,)

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

            output_states += (hidden_states,)

        return hidden_states, output_states


class SimpleCrossAttnDownBlock2D(nn.Layer):
    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,
        attn_num_head_channels=1,
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_downsample=True,
    ):
        super().__init__()

        self.has_cross_attention = True

        resnets = []
        attentions = []

        self.attn_num_head_channels = attn_num_head_channels
        self.num_heads = out_channels // self.attn_num_head_channels

        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,
                )
            )
            attentions.append(
                CrossAttention(
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
                    dim_head=attn_num_head_channels,
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
                    processor=CrossAttnAddedKVProcessor(),
                )
            )
        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

        if add_downsample:
            self.downsamplers = nn.LayerList(
                [
                    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,
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None, cross_attention_kwargs=None
    ):
        output_states = ()
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}

        for resnet, attn in zip(self.resnets, self.attentions):
            # resnet
            hidden_states = resnet(hidden_states, temb)

            # attn
            hidden_states = attn(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=attention_mask,
                **cross_attention_kwargs,
            )

            output_states += (hidden_states,)

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

            output_states += (hidden_states,)

        return hidden_states, output_states


class AttnUpBlock2D(nn.Layer):
    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,
        attn_num_head_channels=1,
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

        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,
                )
            )
            attentions.append(
                AttentionBlock(
                    out_channels,
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
                    norm_num_groups=resnet_groups,
                )
            )

        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

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

    def forward(self, hidden_states, res_hidden_states_tuple, temb=None):
        for resnet, attn in zip(self.resnets, self.attentions):
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = paddle.concat([hidden_states, res_hidden_states], axis=1)

            hidden_states = resnet(hidden_states, temb)
            hidden_states = attn(hidden_states)

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

        return hidden_states


class CrossAttnUpBlock2D(nn.Layer):
    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,
        attn_num_head_channels=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 = []
        attentions = []
        self.has_cross_attention = True
        self.attn_num_head_channels = attn_num_head_channels

        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,
                )
            )
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
                        attn_num_head_channels,
                        out_channels // attn_num_head_channels,
                        in_channels=out_channels,
                        num_layers=1,
                        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,
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
                        attn_num_head_channels,
                        out_channels // attn_num_head_channels,
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
                )
        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

        if add_upsample:
            self.upsamplers = nn.LayerList([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,
        encoder_hidden_states=None,
        cross_attention_kwargs=None,
        upsample_size=None,
        attention_mask=None,
    ):
        # TODO(Patrick, William) - attention mask is not used
        for resnet, attn in zip(self.resnets, self.attentions):
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = paddle.concat([hidden_states, res_hidden_states], axis=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)[0]  # move [0]
                        else:
                            return module(*inputs)

                    return custom_forward

                hidden_states = recompute(create_custom_forward(resnet), hidden_states, temb)
                hidden_states = recompute(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    cross_attention_kwargs,
                )  # [0]
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    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 UpBlock2D(nn.Layer):
    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 = []

        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,
                )
            )

        self.resnets = nn.LayerList(resnets)

        if add_upsample:
            self.upsamplers = nn.LayerList([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):
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = paddle.concat([hidden_states, res_hidden_states], axis=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 = recompute(create_custom_forward(resnet), hidden_states, temb)
            else:
                hidden_states = resnet(hidden_states, temb)

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

        return hidden_states


class UpDecoderBlock2D(nn.Layer):
    def __init__(
        self,
        in_channels: int,
        out_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 = []

        for i in range(num_layers):
            input_channels = in_channels if i == 0 else out_channels

            resnets.append(
                ResnetBlock2D(
                    in_channels=input_channels,
                    out_channels=out_channels,
                    temb_channels=None,
                    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,
                )
            )

        self.resnets = nn.LayerList(resnets)

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

    def forward(self, hidden_states):
        for resnet in self.resnets:
            hidden_states = resnet(hidden_states, temb=None)

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

        return hidden_states


class AttnUpDecoderBlock2D(nn.Layer):
    def __init__(
        self,
        in_channels: int,
        out_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,
        attn_num_head_channels=1,
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

        for i in range(num_layers):
            input_channels = in_channels if i == 0 else out_channels

            resnets.append(
                ResnetBlock2D(
                    in_channels=input_channels,
                    out_channels=out_channels,
                    temb_channels=None,
                    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,
                )
            )
            attentions.append(
                AttentionBlock(
                    out_channels,
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
                    norm_num_groups=resnet_groups,
                )
            )

        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

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

    def forward(self, hidden_states):
        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb=None)
            hidden_states = attn(hidden_states)

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

        return hidden_states


class AttnSkipUpBlock2D(nn.Layer):
    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_pre_norm: bool = True,
        attn_num_head_channels=1,
        output_scale_factor=np.sqrt(2.0),
        upsample_padding=1,
        add_upsample=True,
    ):
        super().__init__()
        self.attentions = nn.LayerList([])
        self.resnets = nn.LayerList([])

        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

            self.resnets.append(
                ResnetBlock2D(
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min(resnet_in_channels + res_skip_channels // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    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,
                )
            )

        self.attentions.append(
            AttentionBlock(
                out_channels,
                num_head_channels=attn_num_head_channels,
                rescale_output_factor=output_scale_factor,
                eps=resnet_eps,
            )
        )

        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
        if add_upsample:
            self.resnet_up = ResnetBlock2D(
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                groups_out=min(out_channels // 4, 32),
                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,
                use_in_shortcut=True,
                up=True,
                kernel="fir",
            )
            self.skip_conv = nn.Conv2D(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
            self.skip_norm = nn.GroupNorm(
                num_groups=min(out_channels // 4, 32), num_channels=out_channels, epsilon=resnet_eps
            )
            self.act = nn.Silu()
        else:
            self.resnet_up = None
            self.skip_conv = None
            self.skip_norm = None
            self.act = None

    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None):
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = paddle.concat([hidden_states, res_hidden_states], axis=1)
            hidden_states = resnet(hidden_states, temb)

        hidden_states = self.attentions[0](hidden_states)

        if skip_sample is not None:
            skip_sample = self.upsampler(skip_sample)
        else:
            skip_sample = 0

        if self.resnet_up is not None:
            skip_sample_states = self.skip_norm(hidden_states)
            skip_sample_states = self.act(skip_sample_states)
            skip_sample_states = self.skip_conv(skip_sample_states)

            skip_sample = skip_sample + skip_sample_states

            hidden_states = self.resnet_up(hidden_states, temb)

        return hidden_states, skip_sample


class SkipUpBlock2D(nn.Layer):
    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_pre_norm: bool = True,
        output_scale_factor=np.sqrt(2.0),
        add_upsample=True,
        upsample_padding=1,
    ):
        super().__init__()
        self.resnets = nn.LayerList([])

        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

            self.resnets.append(
                ResnetBlock2D(
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min((resnet_in_channels + res_skip_channels) // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    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,
                )
            )

        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
        if add_upsample:
            self.resnet_up = ResnetBlock2D(
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                groups_out=min(out_channels // 4, 32),
                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,
                use_in_shortcut=True,
                up=True,
                kernel="fir",
            )
            self.skip_conv = nn.Conv2D(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
            self.skip_norm = nn.GroupNorm(
                num_groups=min(out_channels // 4, 32), num_channels=out_channels, epsilon=resnet_eps
            )
            self.act = nn.Silu()
        else:
            self.resnet_up = None
            self.skip_conv = None
            self.skip_norm = None
            self.act = None

    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None):
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = paddle.concat([hidden_states, res_hidden_states], axis=1)

            hidden_states = resnet(hidden_states, temb)

        if skip_sample is not None:
            skip_sample = self.upsampler(skip_sample)
        else:
            skip_sample = 0

        if self.resnet_up is not None:
            skip_sample_states = self.skip_norm(hidden_states)
            skip_sample_states = self.act(skip_sample_states)
            skip_sample_states = self.skip_conv(skip_sample_states)

            skip_sample = skip_sample + skip_sample_states

            hidden_states = self.resnet_up(hidden_states, temb)

        return hidden_states, skip_sample


class ResnetUpsampleBlock2D(nn.Layer):
    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 = []

        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,
                )
            )

        self.resnets = nn.LayerList(resnets)

        if add_upsample:
            self.upsamplers = nn.LayerList(
                [
                    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,
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = paddle.concat([hidden_states, res_hidden_states], axis=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 = recompute(create_custom_forward(resnet), hidden_states, temb)
            else:
                hidden_states = resnet(hidden_states, temb)

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

        return hidden_states


class SimpleCrossAttnUpBlock2D(nn.Layer):
    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,
        attn_num_head_channels=1,
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
        self.attn_num_head_channels = attn_num_head_channels

        self.num_heads = out_channels // self.attn_num_head_channels

        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,
                )
            )
            attentions.append(
                CrossAttention(
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
                    dim_head=attn_num_head_channels,
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
                    processor=CrossAttnAddedKVProcessor(),
                )
            )
        self.attentions = nn.LayerList(attentions)
        self.resnets = nn.LayerList(resnets)

        if add_upsample:
            self.upsamplers = nn.LayerList(
                [
                    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,
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states,
        res_hidden_states_tuple,
        temb=None,
        encoder_hidden_states=None,
        upsample_size=None,
        attention_mask=None,
        cross_attention_kwargs=None,
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
        for resnet, attn in zip(self.resnets, self.attentions):
            # resnet
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = paddle.concat([hidden_states, res_hidden_states], axis=1)

            hidden_states = resnet(hidden_states, temb)

            # attn
            hidden_states = attn(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=attention_mask,
                **cross_attention_kwargs,
            )

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

        return hidden_states