extensions/diffusers_diffsplat/models/unets/unet_mv2d_blocks.py

from diffusers.models.unets.unet_2d_blocks import *

from extensions.diffusers_diffsplat.models.transformers import TransformerMV2DModel


# Copied from diffusers.models.unets.unet_2d_blocks.get_down_block
# The only modifications: `CrossAttnDownBlock2D` -> `CrossAttnDownBlockMV2D`
def get_down_mvblock(
    down_block_type: str,
    num_layers: int,
    in_channels: int,
    out_channels: int,
    temb_channels: int,
    add_downsample: bool,
    resnet_eps: float,
    resnet_act_fn: str,
    transformer_layers_per_block: int = 1,
    num_attention_heads: Optional[int] = None,
    resnet_groups: Optional[int] = None,
    cross_attention_dim: Optional[int] = None,
    downsample_padding: Optional[int] = None,
    dual_cross_attention: bool = False,
    use_linear_projection: bool = False,
    only_cross_attention: bool = False,
    upcast_attention: bool = False,
    resnet_time_scale_shift: str = "default",
    attention_type: str = "default",
    resnet_skip_time_act: bool = False,
    resnet_out_scale_factor: float = 1.0,
    cross_attention_norm: Optional[str] = None,
    attention_head_dim: Optional[int] = None,
    downsample_type: Optional[str] = None,
    dropout: float = 0.0,
):
    # If attn head dim is not defined, we default it to the number of heads
    if attention_head_dim is None:
        logger.warning(
            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

    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,
            dropout=dropout,
            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,
        )
    # Multi-view Self-Attn
    elif down_block_type == "CrossAttnDownBlockMV2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockMV2D")
        return CrossAttnDownBlockMV2D(
            num_layers=num_layers,
            transformer_layers_per_block=transformer_layers_per_block,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            dropout=dropout,
            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,
            attention_type=attention_type,
        )
    raise ValueError(f"{down_block_type} does not exist.")


# Copied from diffusers.models.unets.unet_2d_blocks.get_mid_block
# The only modifications: `UNetMidBlock2DCrossAttn` -> `UNetMidBlockMV2DCrossAttn`
def get_mid_mvblock(
    mid_block_type: str,
    temb_channels: int,
    in_channels: int,
    resnet_eps: float,
    resnet_act_fn: str,
    resnet_groups: int,
    output_scale_factor: float = 1.0,
    transformer_layers_per_block: int = 1,
    num_attention_heads: Optional[int] = None,
    cross_attention_dim: Optional[int] = None,
    dual_cross_attention: bool = False,
    use_linear_projection: bool = False,
    mid_block_only_cross_attention: bool = False,
    upcast_attention: bool = False,
    resnet_time_scale_shift: str = "default",
    attention_type: str = "default",
    resnet_skip_time_act: bool = False,
    cross_attention_norm: Optional[str] = None,
    attention_head_dim: Optional[int] = 1,
    dropout: float = 0.0,
):
    # Multi-view Self-Attn
    if mid_block_type == "UNetMidBlockMV2DCrossAttn":
        return UNetMidBlockMV2DCrossAttn(
            transformer_layers_per_block=transformer_layers_per_block,
            in_channels=in_channels,
            temb_channels=temb_channels,
            dropout=dropout,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            output_scale_factor=output_scale_factor,
            resnet_time_scale_shift=resnet_time_scale_shift,
            cross_attention_dim=cross_attention_dim,
            num_attention_heads=num_attention_heads,
            resnet_groups=resnet_groups,
            dual_cross_attention=dual_cross_attention,
            use_linear_projection=use_linear_projection,
            upcast_attention=upcast_attention,
            attention_type=attention_type,
        )
    elif mid_block_type is None:
        return None
    else:
        raise ValueError(f"unknown mid_block_type : {mid_block_type}")


# Copied from diffusers.models.unets.unet_2d_blocks.get_up_block
# The only modifications: `CrossAttnUpBlock2D` -> `CrossAttnUpBlockMV2D`
def get_up_mvblock(
    up_block_type: str,
    num_layers: int,
    in_channels: int,
    out_channels: int,
    prev_output_channel: int,
    temb_channels: int,
    add_upsample: bool,
    resnet_eps: float,
    resnet_act_fn: str,
    resolution_idx: Optional[int] = None,
    transformer_layers_per_block: int = 1,
    num_attention_heads: Optional[int] = None,
    resnet_groups: Optional[int] = None,
    cross_attention_dim: Optional[int] = None,
    dual_cross_attention: bool = False,
    use_linear_projection: bool = False,
    only_cross_attention: bool = False,
    upcast_attention: bool = False,
    resnet_time_scale_shift: str = "default",
    attention_type: str = "default",
    resnet_skip_time_act: bool = False,
    resnet_out_scale_factor: float = 1.0,
    cross_attention_norm: Optional[str] = None,
    attention_head_dim: Optional[int] = None,
    upsample_type: Optional[str] = None,
    dropout: float = 0.0,
) -> nn.Module:
    # If attn head dim is not defined, we default it to the number of heads
    if attention_head_dim is None:
        logger.warning(
            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

    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,
            resolution_idx=resolution_idx,
            dropout=dropout,
            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,
        )
    # Multi-view Self-Attn
    elif up_block_type == "CrossAttnUpBlockMV2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockMV2D")
        return CrossAttnUpBlockMV2D(
            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,
            resolution_idx=resolution_idx,
            dropout=dropout,
            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,
            attention_type=attention_type,
        )

    raise ValueError(f"{up_block_type} does not exist.")


# Copied from diffusers.models.unets.unet_2d_blocks.CrossAttnDownBlock2D
# The only modifications: `Transformer2DModel` -> `TransformerMV2DModel`
class CrossAttnDownBlockMV2D(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: Union[int, Tuple[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: int = 1,
        cross_attention_dim: int = 1280,
        output_scale_factor: float = 1.0,
        downsample_padding: int = 1,
        add_downsample: bool = True,
        dual_cross_attention: bool = False,
        use_linear_projection: bool = False,
        only_cross_attention: bool = False,
        upcast_attention: bool = False,
        attention_type: str = "default",
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads
        if isinstance(transformer_layers_per_block, int):
            transformer_layers_per_block = [transformer_layers_per_block] * num_layers

        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(
                    TransformerMV2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=transformer_layers_per_block[i],
                        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,
                        attention_type=attention_type,
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
                        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.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        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.Tensor,
        temb: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        additional_residuals: Optional[torch.Tensor] = None,
    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, ...]]:
        if cross_attention_kwargs is not None:
            if cross_attention_kwargs.get("scale", None) is not None:
                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")

        output_states = ()

        blocks = list(zip(self.resnets, self.attentions))

        for i, (resnet, attn) in enumerate(blocks):
            if torch.is_grad_enabled() 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 = 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]
            else:
                hidden_states = resnet(hidden_states, temb)
                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]

            # apply additional residuals to the output of the last pair of resnet and attention blocks
            if i == len(blocks) - 1 and additional_residuals is not None:
                hidden_states = hidden_states + additional_residuals

            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


# Copied from diffusers.models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn
# The only modifications: `Transformer2DModel` -> `TransformerMV2DModel`
class UNetMidBlockMV2DCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        out_channels: Optional[int] = None,
        dropout: float = 0.0,
        num_layers: int = 1,
        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_groups_out: Optional[int] = None,
        resnet_pre_norm: bool = True,
        num_attention_heads: int = 1,
        output_scale_factor: float = 1.0,
        cross_attention_dim: int = 1280,
        dual_cross_attention: bool = False,
        use_linear_projection: bool = False,
        upcast_attention: bool = False,
        attention_type: str = "default",
    ):
        super().__init__()

        out_channels = out_channels or in_channels
        self.in_channels = in_channels
        self.out_channels = out_channels

        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)

        # support for variable transformer layers per block
        if isinstance(transformer_layers_per_block, int):
            transformer_layers_per_block = [transformer_layers_per_block] * num_layers

        resnet_groups_out = resnet_groups_out or resnet_groups

        # there is always at least one resnet
        resnets = [
            ResnetBlock2D(
                in_channels=in_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=resnet_groups,
                groups_out=resnet_groups_out,
                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 i in range(num_layers):
            if not dual_cross_attention:
                attentions.append(
                    TransformerMV2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=transformer_layers_per_block[i],
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups_out,
                        use_linear_projection=use_linear_projection,
                        upcast_attention=upcast_attention,
                        attention_type=attention_type,
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
                        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,
                    )
                )
            resnets.append(
                ResnetBlock2D(
                    in_channels=out_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups_out,
                    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.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        temb: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        if cross_attention_kwargs is not None:
            if cross_attention_kwargs.get("scale", None) is not None:
                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")

        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            if torch.is_grad_enabled() 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 = 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 = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
            else:
                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 = resnet(hidden_states, temb)

        return hidden_states


# Copied from diffusers.models.unets.unet_2d_blocks.CrossAttnUpBlock2D
# The only modifications: `Transformer2DModel` -> `TransformerMV2DModel`
class CrossAttnUpBlockMV2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
        resolution_idx: Optional[int] = None,
        dropout: float = 0.0,
        num_layers: int = 1,
        transformer_layers_per_block: Union[int, Tuple[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: int = 1,
        cross_attention_dim: int = 1280,
        output_scale_factor: float = 1.0,
        add_upsample: bool = True,
        dual_cross_attention: bool = False,
        use_linear_projection: bool = False,
        only_cross_attention: bool = False,
        upcast_attention: bool = False,
        attention_type: str = "default",
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads

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

        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(
                    TransformerMV2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=transformer_layers_per_block[i],
                        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,
                        attention_type=attention_type,
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
                        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.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        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
        self.resolution_idx = resolution_idx

    def forward(
        self,
        hidden_states: torch.Tensor,
        res_hidden_states_tuple: Tuple[torch.Tensor, ...],
        temb: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        upsample_size: Optional[int] = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        if cross_attention_kwargs is not None:
            if cross_attention_kwargs.get("scale", None) is not None:
                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")

        is_freeu_enabled = (
            getattr(self, "s1", None)
            and getattr(self, "s2", None)
            and getattr(self, "b1", None)
            and getattr(self, "b2", 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]

            # FreeU: Only operate on the first two stages
            if is_freeu_enabled:
                hidden_states, res_hidden_states = apply_freeu(
                    self.resolution_idx,
                    hidden_states,
                    res_hidden_states,
                    s1=self.s1,
                    s2=self.s2,
                    b1=self.b1,
                    b2=self.b2,
                )

            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

            if torch.is_grad_enabled() 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 = 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]
            else:
                hidden_states = resnet(hidden_states, temb)
                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]

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

        return hidden_states