ip_adapter/attention_processor.py

# 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

# Global Variable
global_concept_mask = []
attn_mask_logs = {}
text_attn_map_logs = {}
image_attn_map_logs = {}


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

    def __init__(
        self,
        hidden_size=None,
        cross_attention_dim=None,
    ):
        super().__init__()

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

        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):
        super().__init__()

        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,
        *args,
        **kwargs,
    ):
        global global_concept_mask
        global attn_mask_logs
        global text_attn_map_logs
        global image_attn_map_logs
        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)

        # 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 ConceptrolAttnProcessor(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,
        textual_concept_idxs=None,
        name=None,
        global_masking=False,
        adaptive_scale_mask=False,
        concept_mask_layer=None,
    ):
        super().__init__()

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

        self.textual_concept_idxs = textual_concept_idxs
        self.name = name

        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.global_masking = global_masking
        self.adaptive_scale_mask = adaptive_scale_mask

        if concept_mask_layer is None:
            concept_mask_layer = [
                "mid_block.attentions.0.transformer_blocks.0.attn2.processor"
            ]  # For SD
            print("Warning: Using default concept mask layer for SD. For SDXL, use 'up_blocks.0.attentions.1.transformer_blocks.5.attn2.processor'")
            # concept_mask_layer = ['up_blocks.0.attentions.1.transformer_blocks.1.attn2.processor'] # For SDXL
        self.concept_mask_layer = concept_mask_layer

    def set_global_view(self, attn_procs):
        self.attn_procs = attn_procs
        # print(self.name, self.attn_procs.keys())

    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
        *args,
        **kwargs,
    ):
        global global_concept_mask
        global attn_mask_logs

        if self.textual_concept_idxs is None:
            raise ValueError(
                "textual_concept_idxs should be provided for ConceptrolAttnProcessor"
            )
        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 = 77  # Both SD and SDXL use 77 as length of text tokens
            encoder_hidden_states, ip_hidden_states_cat = (
                encoder_hidden_states[:, :end_pos, :],
                encoder_hidden_states[:, end_pos:, :],
            )
            num_concepts = ip_hidden_states_cat.shape[1] // self.num_tokens
            ip_hidden_states_list = torch.chunk(
                ip_hidden_states_cat, num_concepts, dim=1
            )
            assert len(ip_hidden_states_list) == len(
                self.textual_concept_idxs
            ), f"register_idxs should have the same length as the number of concepts, but got {len(ip_hidden_states_list)} and {len(self.textual_concept_idxs)}"
            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)  # [16, 4096, 40]
        key = attn.head_to_batch_dim(key)  # [16, 77, 40]
        value = attn.head_to_batch_dim(value)  # [16, 77, 40]

        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)

        concept_mask_layer = self.concept_mask_layer
        if len(global_concept_mask) == 0:
            global_concept_mask = [None for _ in range(len(self.textual_concept_idxs))]
        for i in range(len(self.textual_concept_idxs)):
            ip_hidden_states = ip_hidden_states_list[i]
            textual_concept_start_idx, textual_concept_end_idx = (
                self.textual_concept_idxs[i]
            )
            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)  # [16, 4, 40]
            ip_value = attn.head_to_batch_dim(ip_value)  # [16, 4, 40]

            # attention_probs: [20/40, 4096, 77]

            ip_attention_mask = attention_probs[
                :, :, textual_concept_start_idx:textual_concept_end_idx
            ]  # [16, 4096, T]
            ip_attention_mask = torch.mean(
                ip_attention_mask, dim=-1, keepdim=True
            )  # [16, 4096, 1]
            ip_attention_mask = attn.batch_to_head_dim(
                ip_attention_mask
            )  # [2, 4096, 8]
            ip_attention_mask = torch.mean(
                ip_attention_mask, dim=-1, keepdim=True
            )  # [2, 4096, 1]

            ip_attention_mask = ip_attention_mask / (
                torch.amax(ip_attention_mask, dim=-2, keepdim=True) + 1e-6
            )

            ip_attention_mask = ip_attention_mask[1:2]  # (use the classifier one)

            # Visualization
            if self.name not in attn_mask_logs:
                attn_mask_logs[self.name] = []
                text_attn_map_logs[self.name] = []
                image_attn_map_logs[self.name] = []
            attn_mask_logs[self.name].append(
                ip_attention_mask.detach().cpu().numpy()[0, :, 0]
            )
            text_attn_map_logs[self.name].append(
                ip_attention_mask.detach().cpu().numpy()[0, :, 0]
            )

            if self.global_masking and (
                self.name == concept_mask_layer[0]
            ):
                global_concept_mask[i] = ip_attention_mask

            if (
                self.global_masking
                and self.name != concept_mask_layer[0]
                and global_concept_mask[i] is not None
            ):
                original_dim = int(global_concept_mask[i].shape[1] ** 0.5)
                target_dim = int(hidden_states.shape[1] ** 0.5)
                global_concept_mask_2d = global_concept_mask[i].view(
                    global_concept_mask[i].shape[0], 1, original_dim, original_dim
                )
                resized_global_concept_mask_2d = F.interpolate(
                    global_concept_mask_2d,
                    size=(target_dim, target_dim),
                    mode="nearest",
                )
                resized_global_concept_mask = resized_global_concept_mask_2d.view(
                    global_concept_mask[i].shape[0], -1, 1
                )
                ip_attention_mask = resized_global_concept_mask

            ip_attention_probs = attn.get_attention_scores(
                query, ip_key, None
            )  # [16, 4096, 4]

            #  Visualization
            ip_attention_map = attention_probs[:, :, 15:16]  # [16, 4096, T]
            ip_attention_map = torch.mean(
                ip_attention_map, dim=-1, keepdim=True
            )  # [16, 4096, 1]
            ip_attention_map = torch.mean(
                ip_attention_map, dim=-1, keepdim=True
            )  # [16, 4096, 1]
            ip_attention_map = attn.batch_to_head_dim(ip_attention_map)  # [2, 4096, 8]
            ip_attention_map = torch.mean(
                ip_attention_map, dim=-1, keepdim=True
            )  # [2, 4096, 1]
            ip_attention_map = ip_attention_map / (
                torch.amax(ip_attention_map, dim=-2, keepdim=True) + 1e-6
            )
            ip_attention_map = ip_attention_map[1:2]  # (use the classifier one)
            image_attn_map_logs[self.name].append(
                ip_attention_map.detach().cpu().numpy()[0, :, 0]
            )

            ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)  # [16, 4096, 40]
            ip_hidden_states = attn.batch_to_head_dim(
                ip_hidden_states
            )  # [2, 4096, 320]
            ip_hidden_states = ip_hidden_states * ip_attention_mask

            if self.adaptive_scale_mask:
                raise ValueError("adaptive_scale_mask is deprecated already")

            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,
    ):
        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."
            )

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

        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):
        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,
        *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
        )

        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)

        # 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


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

    def __init__(self, num_tokens=4):
        self.num_tokens = num_tokens

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

        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):
        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

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

        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