# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
import torch
import torch.nn as nn
import torch.nn.functional as F


class AttnProcessor(nn.Module):
    r"""
    Default processor for performing attention-related computations.
    """

    def __init__(
        self,
        hidden_size=None,
        cross_attention_dim=None,
        save_in_unet='down',
        atten_control=None,
    ):
        super().__init__()
        self.atten_control = atten_control
        self.save_in_unet = save_in_unet

    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
    ):
        residual = hidden_states

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )
        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        query = attn.head_to_batch_dim(query)
        key = attn.head_to_batch_dim(key)
        value = attn.head_to_batch_dim(value)

        attention_probs = attn.get_attention_scores(query, key, attention_mask)
        hidden_states = torch.bmm(attention_probs, value)
        hidden_states = attn.batch_to_head_dim(hidden_states)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class IPAttnProcessor(nn.Module):
    r"""
    Attention processor for IP-Adapater.
    Args:
        hidden_size (`int`):
            The hidden size of the attention layer.
        cross_attention_dim (`int`):
            The number of channels in the `encoder_hidden_states`.
        scale (`float`, defaults to 1.0):
            the weight scale of image prompt.
        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
            The context length of the image features.
    """

    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4, skip=False,save_in_unet='down', atten_control=None):
        super().__init__()

        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.scale = scale
        self.num_tokens = num_tokens
        self.skip = skip

        self.atten_control = atten_control
        self.save_in_unet = save_in_unet

        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)

    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
    ):
        residual = hidden_states

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )
        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        else:
            # get encoder_hidden_states, ip_hidden_states
            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
            encoder_hidden_states, ip_hidden_states = (
                encoder_hidden_states[:, :end_pos, :],
                encoder_hidden_states[:, end_pos:, :],
            )
            if attn.norm_cross:
                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        query = attn.head_to_batch_dim(query)
        key = attn.head_to_batch_dim(key)
        value = attn.head_to_batch_dim(value)

        attention_probs = attn.get_attention_scores(query, key, attention_mask)
        hidden_states = torch.bmm(attention_probs, value)
        hidden_states = attn.batch_to_head_dim(hidden_states)

        if not self.skip:
            # for ip-adapter
            ip_key = self.to_k_ip(ip_hidden_states)
            ip_value = self.to_v_ip(ip_hidden_states)

            ip_key = attn.head_to_batch_dim(ip_key)
            ip_value = attn.head_to_batch_dim(ip_value)

            ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
            self.attn_map = ip_attention_probs
            ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
            ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)

            hidden_states = hidden_states + self.scale * ip_hidden_states

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class AttnProcessor2_0(torch.nn.Module):
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
    """

    def __init__(
        self,
        hidden_size=None,
        cross_attention_dim=None,
        save_in_unet='down',
            atten_control=None,
    ):
        super().__init__()
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
        self.atten_control = atten_control
        self.save_in_unet = save_in_unet

    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
    ):
        residual = hidden_states

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )

        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        # the output of sdp = (batch, num_heads, seq_len, head_dim)
        # TODO: add support for attn.scale when we move to Torch 2.1
        hidden_states = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
        )

        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states = hidden_states.to(query.dtype)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class IPAttnProcessor2_0(torch.nn.Module):
    r"""
    Attention processor for IP-Adapater for PyTorch 2.0.
    Args:
        hidden_size (`int`):
            The hidden size of the attention layer.
        cross_attention_dim (`int`):
            The number of channels in the `encoder_hidden_states`.
        scale (`float`, defaults to 1.0):
            the weight scale of image prompt.
        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
            The context length of the image features.
    """

    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4, skip=False,save_in_unet='down', atten_control=None):
        super().__init__()

        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.scale = scale
        self.num_tokens = num_tokens
        self.skip = skip

        self.atten_control = atten_control
        self.save_in_unet = save_in_unet

        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)

    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
    ):
        residual = hidden_states

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )

        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        else:
            # get encoder_hidden_states, ip_hidden_states
            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
            encoder_hidden_states, ip_hidden_states = (
                encoder_hidden_states[:, :end_pos, :],
                encoder_hidden_states[:, end_pos:, :],
            )
            if attn.norm_cross:
                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        # the output of sdp = (batch, num_heads, seq_len, head_dim)
        # TODO: add support for attn.scale when we move to Torch 2.1
        hidden_states = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
        )

        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states = hidden_states.to(query.dtype)

        if not self.skip:
            # for ip-adapter
            ip_key = self.to_k_ip(ip_hidden_states)
            ip_value = self.to_v_ip(ip_hidden_states)

            ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
            ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

            # the output of sdp = (batch, num_heads, seq_len, head_dim)
            # TODO: add support for attn.scale when we move to Torch 2.1
            ip_hidden_states = F.scaled_dot_product_attention(
                query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
            )
            with torch.no_grad():
                self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
                #print(self.attn_map.shape)

            ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
            ip_hidden_states = ip_hidden_states.to(query.dtype)

            hidden_states = hidden_states + self.scale * ip_hidden_states

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class IP_CS_AttnProcessor2_0(torch.nn.Module):
    r"""
    Attention processor for IP-Adapater for PyTorch 2.0.
    Args:
        hidden_size (`int`):
            The hidden size of the attention layer.
        cross_attention_dim (`int`):
            The number of channels in the `encoder_hidden_states`.
        scale (`float`, defaults to 1.0):
            the weight scale of image prompt.
        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
            The context length of the image features.
    """

    def __init__(self, hidden_size, cross_attention_dim=None, content_scale=1.0,style_scale=1.0, num_content_tokens=4,num_style_tokens=4,
                 skip=False,content=False, style=False):
        super().__init__()

        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.content_scale = content_scale
        self.style_scale = style_scale
        self.num_content_tokens = num_content_tokens
        self.num_style_tokens = num_style_tokens
        self.skip = skip

        self.content = content
        self.style = style

        if self.content or self.style:
            self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
            self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
        self.to_k_ip_content =None
        self.to_v_ip_content =None

    def set_content_ipa(self,content_scale=1.0):

        self.to_k_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
        self.to_v_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
        self.content_scale=content_scale
        self.content =True

    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
    ):
        residual = hidden_states

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )

        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        else:
            # get encoder_hidden_states, ip_hidden_states
            end_pos = encoder_hidden_states.shape[1] - self.num_content_tokens-self.num_style_tokens
            encoder_hidden_states, ip_content_hidden_states,ip_style_hidden_states = (
                encoder_hidden_states[:, :end_pos, :],
                encoder_hidden_states[:, end_pos:end_pos + self.num_content_tokens, :],
                encoder_hidden_states[:, end_pos + self.num_content_tokens:, :],
            )
            if attn.norm_cross:
                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        # the output of sdp = (batch, num_heads, seq_len, head_dim)
        # TODO: add support for attn.scale when we move to Torch 2.1
        hidden_states = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
        )

        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states = hidden_states.to(query.dtype)

        if not self.skip and self.content is True:
            # print('content#####################################################')
            # for ip-content-adapter
            if self.to_k_ip_content is None:

                ip_content_key = self.to_k_ip(ip_content_hidden_states)
                ip_content_value = self.to_v_ip(ip_content_hidden_states)
            else:
                ip_content_key = self.to_k_ip_content(ip_content_hidden_states)
                ip_content_value = self.to_v_ip_content(ip_content_hidden_states)

            ip_content_key = ip_content_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
            ip_content_value = ip_content_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

            # the output of sdp = (batch, num_heads, seq_len, head_dim)
            # TODO: add support for attn.scale when we move to Torch 2.1
            ip_content_hidden_states = F.scaled_dot_product_attention(
                query, ip_content_key, ip_content_value, attn_mask=None, dropout_p=0.0, is_causal=False
            )


            ip_content_hidden_states = ip_content_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
            ip_content_hidden_states = ip_content_hidden_states.to(query.dtype)


            hidden_states = hidden_states + self.content_scale * ip_content_hidden_states

        if not self.skip and self.style is True:
            # for ip-style-adapter
            ip_style_key = self.to_k_ip(ip_style_hidden_states)
            ip_style_value = self.to_v_ip(ip_style_hidden_states)

            ip_style_key = ip_style_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
            ip_style_value = ip_style_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

            # the output of sdp = (batch, num_heads, seq_len, head_dim)
            # TODO: add support for attn.scale when we move to Torch 2.1
            ip_style_hidden_states = F.scaled_dot_product_attention(
                query, ip_style_key, ip_style_value, attn_mask=None, dropout_p=0.0, is_causal=False
            )

            ip_style_hidden_states = ip_style_hidden_states.transpose(1, 2).reshape(batch_size, -1,
                                                                                    attn.heads * head_dim)
            ip_style_hidden_states = ip_style_hidden_states.to(query.dtype)

            hidden_states = hidden_states + self.style_scale * ip_style_hidden_states

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states

## for controlnet
class CNAttnProcessor:
    r"""
    Default processor for performing attention-related computations.
    """

    def __init__(self, num_tokens=4,save_in_unet='down',atten_control=None):
        self.num_tokens = num_tokens
        self.atten_control = atten_control
        self.save_in_unet = save_in_unet

    def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None):
        residual = hidden_states

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )
        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        else:
            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
            if attn.norm_cross:
                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        query = attn.head_to_batch_dim(query)
        key = attn.head_to_batch_dim(key)
        value = attn.head_to_batch_dim(value)

        attention_probs = attn.get_attention_scores(query, key, attention_mask)
        hidden_states = torch.bmm(attention_probs, value)
        hidden_states = attn.batch_to_head_dim(hidden_states)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class CNAttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
    """

    def __init__(self, num_tokens=4, save_in_unet='down', atten_control=None):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
        self.num_tokens = num_tokens
        self.atten_control = atten_control
        self.save_in_unet = save_in_unet

    def __call__(
            self,
            attn,
            hidden_states,
            encoder_hidden_states=None,
            attention_mask=None,
            temb=None,
    ):
        residual = hidden_states

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )

        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        else:
            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
            if attn.norm_cross:
                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        # the output of sdp = (batch, num_heads, seq_len, head_dim)
        # TODO: add support for attn.scale when we move to Torch 2.1
        hidden_states = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
        )

        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states = hidden_states.to(query.dtype)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states