from typing import Callable, Optional, Union

import torch
import torch.nn.functional as F
from torch import nn
from diffusers.utils import USE_PEFT_BACKEND
from diffusers.models.lora import LoRALinearLayer





class CacheAttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
    """

    def __init__(self):
        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.cache = {}  # cache hidden states

    def __call__(
            self,
            attn,
            hidden_states: torch.FloatTensor,
            encoder_hidden_states: Optional[torch.FloatTensor] = None,
            attention_mask: Optional[torch.FloatTensor] = None,
            temb: Optional[torch.FloatTensor] = None,
            scale: float = 1.0,
    ) -> torch.FloatTensor:

        self.cache["hidden_states"] = hidden_states  # cache hidden states

        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)

        args = () if USE_PEFT_BACKEND else (scale,)
        query = attn.to_q(hidden_states, *args)

        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, *args)
        value = attn.to_v(encoder_hidden_states, *args)

        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, *args)
        # 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 SAttnProcessor2_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, name, hidden_size, cross_attention_dim=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.")

        super().__init__()

        self.name = name
        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim

    def __call__(
            self,
            attn,
            hidden_states: torch.FloatTensor,
            encoder_hidden_states: Optional[torch.FloatTensor] = None,
            attention_mask: Optional[torch.FloatTensor] = None,
            temb: Optional[torch.FloatTensor] = None,
            scale: float = 1.0,
            cond_hidden_states=None,
            sa_hidden_states=None,
    ) -> torch.FloatTensor:
        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)

        args = () if USE_PEFT_BACKEND else (scale,)
        query = attn.to_q(hidden_states, *args)

        if encoder_hidden_states is None:
            # for reference adapter
            if sa_hidden_states is not None:
                ref_hidden_states = sa_hidden_states[self.name]
                encoder_hidden_states = torch.cat([hidden_states, ref_hidden_states], dim=1)
            else:
                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, *args)
        value = attn.to_v(encoder_hidden_states, *args)

        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, *args)
        # 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 CAttnProcessor2_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, name, hidden_size, cross_attention_dim=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.")

        super().__init__()

        self.name = name
        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim

        # 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: torch.FloatTensor,
            encoder_hidden_states: Optional[torch.FloatTensor] = None,
            attention_mask: Optional[torch.FloatTensor] = None,
            temb: Optional[torch.FloatTensor] = None,
            scale: float = 1.0,
            cond_hidden_states=None,
            sa_hidden_states=None,
    ) -> torch.FloatTensor:
        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)

        args = () if USE_PEFT_BACKEND else (scale,)
        query = attn.to_q(hidden_states, *args)

        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, *args)
        value = attn.to_v(encoder_hidden_states, *args)

        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)

        # for ip
        # if cond_hidden_states:
        # ip_hidden_states = cond_hidden_states
        # 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=attention_mask, dropout_p=0.0, is_causal=False
        # )
        # 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 + ip_hidden_states

        # linear proj
        hidden_states = attn.to_out[0](hidden_states, *args)
        # 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 RefLoraSAttnProcessor2_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, name, hidden_size, cross_attention_dim=None, scale=1.0, rank=128, network_alpha=None, lora_scale=1.0,):
        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.")

        super().__init__()

        self.name = name
        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.to_k_ref = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
        self.to_v_ref = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
        self.scale = scale

        self.rank = rank
        self.lora_scale = lora_scale
        self.to_q_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
        self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
        self.to_v_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
        self.to_out_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
        
    def __call__(
            self,
            attn,
            hidden_states: torch.FloatTensor,
            encoder_hidden_states: Optional[torch.FloatTensor] = None,
            attention_mask: Optional[torch.FloatTensor] = None,
            temb: Optional[torch.FloatTensor] = None,
            scale: float = 1.0,
            num_images_per_prompt=1,
            cond_hidden_states=None,
            sa_hidden_states=None,

    ) -> torch.FloatTensor:
        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)

        args = () if USE_PEFT_BACKEND else (scale,)
        query = attn.to_q(hidden_states, *args) + self.lora_scale * self.to_q_lora(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, *args)  + self.lora_scale * self.to_k_lora(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states, *args)  + self.lora_scale * self.to_v_lora(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)

        # for ref adapter
        if sa_hidden_states is not None:
            ref_hidden_states = sa_hidden_states[self.name]
            # for ref
            ref_key = self.to_k_ref(ref_hidden_states)
            ref_value = self.to_v_ref(ref_hidden_states)
            ref_key = ref_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
            ref_value = ref_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
            ref_hidden_states = F.scaled_dot_product_attention(
                query, ref_key, ref_value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
            )
            ref_hidden_states = ref_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
            ref_hidden_states = ref_hidden_states.to(query.dtype)
            hidden_states = hidden_states + ref_hidden_states * self.scale

        # linear proj
        hidden_states = attn.to_out[0](hidden_states, *args) + self.lora_scale * self.to_out_lora(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 RefSAttnProcessor2_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, name, hidden_size, cross_attention_dim=None, scale=1.0):
        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.")

        super().__init__()

        self.name = name
        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.to_k_ref = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
        self.to_v_ref = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
        self.scale = scale

    def __call__(
            self,
            attn,
            hidden_states: torch.FloatTensor,
            encoder_hidden_states: Optional[torch.FloatTensor] = None,
            attention_mask: Optional[torch.FloatTensor] = None,
            temb: Optional[torch.FloatTensor] = None,
            scale: float = 1.0,
            num_images_per_prompt=1,
            cond_hidden_states=None,
            sa_hidden_states=None,

    ) -> torch.FloatTensor:
        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)

        args = () if USE_PEFT_BACKEND else (scale,)
        query = attn.to_q(hidden_states, *args)

        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, *args)
        value = attn.to_v(encoder_hidden_states, *args)

        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)

        # for ref adapter
        if sa_hidden_states is not None:
            ref_hidden_states = sa_hidden_states[self.name]
            # for ref
            ref_key = self.to_k_ref(ref_hidden_states)
            ref_value = self.to_v_ref(ref_hidden_states)
            ref_key = ref_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
            ref_value = ref_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
            ref_hidden_states = F.scaled_dot_product_attention(
                query, ref_key, ref_value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
            )
            ref_hidden_states = ref_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
            ref_hidden_states = ref_hidden_states.to(query.dtype)
            hidden_states = hidden_states + ref_hidden_states * self.scale

        # linear proj
        hidden_states = attn.to_out[0](hidden_states, *args)
        # 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):
        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.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,
            sa_hidden_states=None,
            scale: float = 1.0,
    ):
        # attn原始的attn模块
        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:
            if sa_hidden_states is not None:
                ref_hidden_states = sa_hidden_states[self.name]
                # print(ref_hidden_states.shape, hidden_states.shape)
                encoder_hidden_states = torch.cat([hidden_states, ref_hidden_states], dim=1)
            else:
                encoder_hidden_states = hidden_states
        else:
            # get encoder_hidden_states, ip_hidden_states
            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
            if end_pos != 89:
                encoder_hidden_states = encoder_hidden_states
                ip_hidden_states = None
            else:
                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)

        # make sure the ipa is in the inference stage
        if ip_hidden_states is not None:
            # 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 LoRAIPAttnProcessor2_0(nn.Module):
    r"""
    Processor for implementing the LoRA attention mechanism.

    Args:
        hidden_size (`int`, *optional*):
            The hidden size of the attention layer.
        cross_attention_dim (`int`, *optional*):
            The number of channels in the `encoder_hidden_states`.
        rank (`int`, defaults to 4):
            The dimension of the LoRA update matrices.
        network_alpha (`int`, *optional*):
            Equivalent to `alpha` but it's usage is specific to Kohya (A1111) style LoRAs.
    """

    def __init__(self, hidden_size, cross_attention_dim=None, rank=128, network_alpha=None, lora_scale=1.0, scale=1.0,
                 num_tokens=4):
        super().__init__()

        self.rank = rank
        self.lora_scale = lora_scale
        self.num_tokens = num_tokens

        self.to_q_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
        self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
        self.to_v_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
        self.to_out_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)

        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.scale = scale

        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, scale=1.0, temb=None, *args,
            **kwargs,
    ):
        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) + self.lora_scale * self.to_q_lora(hidden_states)
        # query = attn.head_to_batch_dim(query)

        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)

        # for text
        key = attn.to_k(encoder_hidden_states) + self.lora_scale * self.to_k_lora(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states) + self.lora_scale * self.to_v_lora(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)

        # for ip
        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
        )

        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) + self.lora_scale * self.to_out_lora(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