models/attention.py

# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


# Some modifications are reimplemented in public environments by Xiao Fu and Mu Hu


from typing import Any, Dict, Optional

import torch
import torch.nn.functional as F
from torch import nn
import xformers

from diffusers.utils import USE_PEFT_BACKEND
from diffusers.utils.torch_utils import maybe_allow_in_graph
from diffusers.models.activations import GEGLU, GELU, ApproximateGELU
from diffusers.models.attention_processor import Attention
from diffusers.models.embeddings import SinusoidalPositionalEmbedding
from diffusers.models.lora import LoRACompatibleLinear
from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm


def _chunked_feed_forward(
    ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None
):
    # "feed_forward_chunk_size" can be used to save memory
    if hidden_states.shape[chunk_dim] % chunk_size != 0:
        raise ValueError(
            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
        )

    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
    if lora_scale is None:
        ff_output = torch.cat(
            [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
            dim=chunk_dim,
        )
    else:
        # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete
        ff_output = torch.cat(
            [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
            dim=chunk_dim,
        )

    return ff_output


@maybe_allow_in_graph
class GatedSelfAttentionDense(nn.Module):
    r"""
    A gated self-attention dense layer that combines visual features and object features.

    Parameters:
        query_dim (`int`): The number of channels in the query.
        context_dim (`int`): The number of channels in the context.
        n_heads (`int`): The number of heads to use for attention.
        d_head (`int`): The number of channels in each head.
    """

    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
        super().__init__()

        # we need a linear projection since we need cat visual feature and obj feature
        self.linear = nn.Linear(context_dim, query_dim)

        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
        self.ff = FeedForward(query_dim, activation_fn="geglu")

        self.norm1 = nn.LayerNorm(query_dim)
        self.norm2 = nn.LayerNorm(query_dim)

        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))

        self.enabled = True

    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
        if not self.enabled:
            return x

        n_visual = x.shape[1]
        objs = self.linear(objs)

        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))

        return x


@maybe_allow_in_graph
class BasicTransformerBlock(nn.Module):
    r"""
    A basic Transformer block.

    Parameters:
        dim (`int`): The number of channels in the input and output.
        num_attention_heads (`int`): The number of heads to use for multi-head attention.
        attention_head_dim (`int`): The number of channels in each head.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
        num_embeds_ada_norm (:
            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
        attention_bias (:
            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
        only_cross_attention (`bool`, *optional*):
            Whether to use only cross-attention layers. In this case two cross attention layers are used.
        double_self_attention (`bool`, *optional*):
            Whether to use two self-attention layers. In this case no cross attention layers are used.
        upcast_attention (`bool`, *optional*):
            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
            Whether to use learnable elementwise affine parameters for normalization.
        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
        final_dropout (`bool` *optional*, defaults to False):
            Whether to apply a final dropout after the last feed-forward layer.
        attention_type (`str`, *optional*, defaults to `"default"`):
            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
        positional_embeddings (`str`, *optional*, defaults to `None`):
            The type of positional embeddings to apply to.
        num_positional_embeddings (`int`, *optional*, defaults to `None`):
            The maximum number of positional embeddings to apply.
    """

    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        dropout=0.0,
        cross_attention_dim: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        attention_bias: bool = False,
        only_cross_attention: bool = False,
        double_self_attention: bool = False,
        upcast_attention: bool = False,
        norm_elementwise_affine: bool = True,
        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
        norm_eps: float = 1e-5,
        final_dropout: bool = False,
        attention_type: str = "default",
        positional_embeddings: Optional[str] = None,
        num_positional_embeddings: Optional[int] = None,
        ada_norm_continous_conditioning_embedding_dim: Optional[int] = None,
        ada_norm_bias: Optional[int] = None,
        ff_inner_dim: Optional[int] = None,
        ff_bias: bool = True,
        attention_out_bias: bool = True,
    ):
        super().__init__()
        self.only_cross_attention = only_cross_attention

        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
        self.use_layer_norm = norm_type == "layer_norm"
        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"

        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
            raise ValueError(
                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
            )

        if positional_embeddings and (num_positional_embeddings is None):
            raise ValueError(
                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
            )

        if positional_embeddings == "sinusoidal":
            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
        else:
            self.pos_embed = None

        # Define 3 blocks. Each block has its own normalization layer.
        # 1. Self-Attn
        if self.use_ada_layer_norm:
            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
        elif self.use_ada_layer_norm_zero:
            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
        elif self.use_ada_layer_norm_continuous:
            self.norm1 = AdaLayerNormContinuous(
                dim,
                ada_norm_continous_conditioning_embedding_dim,
                norm_elementwise_affine,
                norm_eps,
                ada_norm_bias,
                "rms_norm",
            )
        else:
            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)


        self.attn1 = CustomJointAttention(
            query_dim=dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
            upcast_attention=upcast_attention,
            out_bias=attention_out_bias
        )

        # 2. Cross-Attn
        if cross_attention_dim is not None or double_self_attention:
            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
            # the second cross attention block.

            if self.use_ada_layer_norm:
                self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
            elif self.use_ada_layer_norm_continuous:
                self.norm2 = AdaLayerNormContinuous(
                    dim,
                    ada_norm_continous_conditioning_embedding_dim,
                    norm_elementwise_affine,
                    norm_eps,
                    ada_norm_bias,
                    "rms_norm",
                )
            else:
                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)

            self.attn2 = Attention(
                query_dim=dim,
                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                upcast_attention=upcast_attention,
                out_bias=attention_out_bias,
            )  # is self-attn if encoder_hidden_states is none
        else:
            self.norm2 = None
            self.attn2 = None

        # 3. Feed-forward
        if self.use_ada_layer_norm_continuous:
            self.norm3 = AdaLayerNormContinuous(
                dim,
                ada_norm_continous_conditioning_embedding_dim,
                norm_elementwise_affine,
                norm_eps,
                ada_norm_bias,
                "layer_norm",
            )
        elif not self.use_ada_layer_norm_single:
            self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)

        self.ff = FeedForward(
            dim,
            dropout=dropout,
            activation_fn=activation_fn,
            final_dropout=final_dropout,
            inner_dim=ff_inner_dim,
            bias=ff_bias,
        )

        # 4. Fuser
        if attention_type == "gated" or attention_type == "gated-text-image":
            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)

        # 5. Scale-shift for PixArt-Alpha.
        if self.use_ada_layer_norm_single:
            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)

        # let chunk size default to None
        self._chunk_size = None
        self._chunk_dim = 0

    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
        # Sets chunk feed-forward
        self._chunk_size = chunk_size
        self._chunk_dim = dim

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        timestep: Optional[torch.LongTensor] = None,
        cross_attention_kwargs: Dict[str, Any] = None,
        class_labels: Optional[torch.LongTensor] = None,
        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
    ) -> torch.FloatTensor:
        # Notice that normalization is always applied before the real computation in the following blocks.

        # 0. Self-Attention
        batch_size = hidden_states.shape[0]

        if self.use_ada_layer_norm:
            norm_hidden_states = self.norm1(hidden_states, timestep)
        elif self.use_ada_layer_norm_zero:
            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
            )
        elif self.use_layer_norm:
            norm_hidden_states = self.norm1(hidden_states)
        elif self.use_ada_layer_norm_continuous:
            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
        elif self.use_ada_layer_norm_single:
            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
            ).chunk(6, dim=1)
            norm_hidden_states = self.norm1(hidden_states)
            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
            norm_hidden_states = norm_hidden_states.squeeze(1)
        else:
            raise ValueError("Incorrect norm used")

        if self.pos_embed is not None:
            norm_hidden_states = self.pos_embed(norm_hidden_states)

        # 1. Retrieve lora scale.
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0

        # 2. Prepare GLIGEN inputs
        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)

        attn_output = self.attn1(
            norm_hidden_states,
            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
            attention_mask=attention_mask,
            **cross_attention_kwargs,
        )
        if self.use_ada_layer_norm_zero:
            attn_output = gate_msa.unsqueeze(1) * attn_output
        elif self.use_ada_layer_norm_single:
            attn_output = gate_msa * attn_output

        hidden_states = attn_output + hidden_states
        if hidden_states.ndim == 4:
            hidden_states = hidden_states.squeeze(1)

        # 2.5 GLIGEN Control
        if gligen_kwargs is not None:
            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])

        # 3. Cross-Attention
        if self.attn2 is not None:
            if self.use_ada_layer_norm:
                norm_hidden_states = self.norm2(hidden_states, timestep)
            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
                norm_hidden_states = self.norm2(hidden_states)
            elif self.use_ada_layer_norm_single:
                # For PixArt norm2 isn't applied here:
                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
                norm_hidden_states = hidden_states
            elif self.use_ada_layer_norm_continuous:
                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
            else:
                raise ValueError("Incorrect norm")

            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
                norm_hidden_states = self.pos_embed(norm_hidden_states)

            attn_output = self.attn2(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                **cross_attention_kwargs,
            )
            hidden_states = attn_output + hidden_states

        # 4. Feed-forward
        if self.use_ada_layer_norm_continuous:
            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
        elif not self.use_ada_layer_norm_single:
            norm_hidden_states = self.norm3(hidden_states)

        if self.use_ada_layer_norm_zero:
            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]

        if self.use_ada_layer_norm_single:
            norm_hidden_states = self.norm2(hidden_states)
            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp

        if self._chunk_size is not None:
            # "feed_forward_chunk_size" can be used to save memory
            ff_output = _chunked_feed_forward(
                self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale
            )
        else:
            ff_output = self.ff(norm_hidden_states, scale=lora_scale)

        if self.use_ada_layer_norm_zero:
            ff_output = gate_mlp.unsqueeze(1) * ff_output
        elif self.use_ada_layer_norm_single:
            ff_output = gate_mlp * ff_output

        hidden_states = ff_output + hidden_states
        if hidden_states.ndim == 4:
            hidden_states = hidden_states.squeeze(1)

        return hidden_states


class CustomJointAttention(Attention):
    def set_use_memory_efficient_attention_xformers(
        self, use_memory_efficient_attention_xformers: bool, *args, **kwargs
    ):
        processor = XFormersJointAttnProcessor()
        self.set_processor(processor)
        # print("using xformers attention processor")


class XFormersJointAttnProcessor:
    r"""
    Default processor for performing attention-related computations.
    """

    def __call__(
        self,
        attn: Attention,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        temb=None,
        num_tasks=2
    ):
    
        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)

        # from yuancheng; here attention_mask is None
        if attention_mask is not None:
            # expand our mask's singleton query_tokens dimension:
            #   [batch*heads,            1, key_tokens] ->
            #   [batch*heads, query_tokens, key_tokens]
            # so that it can be added as a bias onto the attention scores that xformers computes:
            #   [batch*heads, query_tokens, key_tokens]
            # we do this explicitly because xformers doesn't broadcast the singleton dimension for us.
            _, query_tokens, _ = hidden_states.shape
            attention_mask = attention_mask.expand(-1, query_tokens, -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)

        assert num_tasks == 2  # only support two tasks now

        key_0, key_1 = torch.chunk(key, dim=0, chunks=2)  # keys shape (b t) d c
        value_0, value_1 = torch.chunk(value, dim=0, chunks=2)

        # key = torch.cat([key_1, key_0], dim=0)
        # value = torch.cat([value_1, value_0], dim=0)

        key = torch.cat([key_0, key_1], dim=1)  # (b t) 2d c
        value = torch.cat([value_0, value_1], dim=1)  # (b t) 2d c
        key = torch.cat([key]*2, dim=0)   # (2 b t) 2d c
        value = torch.cat([value]*2, dim=0)  # (2 b t) 2d c

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

        hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
        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


@maybe_allow_in_graph
class TemporalBasicTransformerBlock(nn.Module):
    r"""
    A basic Transformer block for video like data.

    Parameters:
        dim (`int`): The number of channels in the input and output.
        time_mix_inner_dim (`int`): The number of channels for temporal attention.
        num_attention_heads (`int`): The number of heads to use for multi-head attention.
        attention_head_dim (`int`): The number of channels in each head.
        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
    """

    def __init__(
        self,
        dim: int,
        time_mix_inner_dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        cross_attention_dim: Optional[int] = None,
    ):
        super().__init__()
        self.is_res = dim == time_mix_inner_dim

        self.norm_in = nn.LayerNorm(dim)

        # Define 3 blocks. Each block has its own normalization layer.
        # 1. Self-Attn
        self.norm_in = nn.LayerNorm(dim)
        self.ff_in = FeedForward(
            dim,
            dim_out=time_mix_inner_dim,
            activation_fn="geglu",
        )

        self.norm1 = nn.LayerNorm(time_mix_inner_dim)
        self.attn1 = Attention(
            query_dim=time_mix_inner_dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            cross_attention_dim=None,
        )

        # 2. Cross-Attn
        if cross_attention_dim is not None:
            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
            # the second cross attention block.
            self.norm2 = nn.LayerNorm(time_mix_inner_dim)
            self.attn2 = Attention(
                query_dim=time_mix_inner_dim,
                cross_attention_dim=cross_attention_dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
            )  # is self-attn if encoder_hidden_states is none
        else:
            self.norm2 = None
            self.attn2 = None

        # 3. Feed-forward
        self.norm3 = nn.LayerNorm(time_mix_inner_dim)
        self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu")

        # let chunk size default to None
        self._chunk_size = None
        self._chunk_dim = None

    def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs):
        # Sets chunk feed-forward
        self._chunk_size = chunk_size
        # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off
        self._chunk_dim = 1

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        num_frames: int,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
    ) -> torch.FloatTensor:
        # Notice that normalization is always applied before the real computation in the following blocks.
        # 0. Self-Attention
        batch_size = hidden_states.shape[0]

        batch_frames, seq_length, channels = hidden_states.shape
        batch_size = batch_frames // num_frames

        hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels)
        hidden_states = hidden_states.permute(0, 2, 1, 3)
        hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels)

        residual = hidden_states
        hidden_states = self.norm_in(hidden_states)

        if self._chunk_size is not None:
            hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size)
        else:
            hidden_states = self.ff_in(hidden_states)

        if self.is_res:
            hidden_states = hidden_states + residual

        norm_hidden_states = self.norm1(hidden_states)
        attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
        hidden_states = attn_output + hidden_states

        # 3. Cross-Attention
        if self.attn2 is not None:
            norm_hidden_states = self.norm2(hidden_states)
            attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
            hidden_states = attn_output + hidden_states

        # 4. Feed-forward
        norm_hidden_states = self.norm3(hidden_states)

        if self._chunk_size is not None:
            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
        else:
            ff_output = self.ff(norm_hidden_states)

        if self.is_res:
            hidden_states = ff_output + hidden_states
        else:
            hidden_states = ff_output

        hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels)
        hidden_states = hidden_states.permute(0, 2, 1, 3)
        hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels)

        return hidden_states


class SkipFFTransformerBlock(nn.Module):
    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        kv_input_dim: int,
        kv_input_dim_proj_use_bias: bool,
        dropout=0.0,
        cross_attention_dim: Optional[int] = None,
        attention_bias: bool = False,
        attention_out_bias: bool = True,
    ):
        super().__init__()
        if kv_input_dim != dim:
            self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias)
        else:
            self.kv_mapper = None

        self.norm1 = RMSNorm(dim, 1e-06)

        self.attn1 = Attention(
            query_dim=dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
            cross_attention_dim=cross_attention_dim,
            out_bias=attention_out_bias,
        )

        self.norm2 = RMSNorm(dim, 1e-06)

        self.attn2 = Attention(
            query_dim=dim,
            cross_attention_dim=cross_attention_dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
            out_bias=attention_out_bias,
        )

    def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs):
        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}

        if self.kv_mapper is not None:
            encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states))

        norm_hidden_states = self.norm1(hidden_states)

        attn_output = self.attn1(
            norm_hidden_states,
            encoder_hidden_states=encoder_hidden_states,
            **cross_attention_kwargs,
        )

        hidden_states = attn_output + hidden_states

        norm_hidden_states = self.norm2(hidden_states)

        attn_output = self.attn2(
            norm_hidden_states,
            encoder_hidden_states=encoder_hidden_states,
            **cross_attention_kwargs,
        )

        hidden_states = attn_output + hidden_states

        return hidden_states


class FeedForward(nn.Module):
    r"""
    A feed-forward layer.

    Parameters:
        dim (`int`): The number of channels in the input.
        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
    """

    def __init__(
        self,
        dim: int,
        dim_out: Optional[int] = None,
        mult: int = 4,
        dropout: float = 0.0,
        activation_fn: str = "geglu",
        final_dropout: bool = False,
        inner_dim=None,
        bias: bool = True,
    ):
        super().__init__()
        if inner_dim is None:
            inner_dim = int(dim * mult)
        dim_out = dim_out if dim_out is not None else dim
        linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear

        if activation_fn == "gelu":
            act_fn = GELU(dim, inner_dim, bias=bias)
        if activation_fn == "gelu-approximate":
            act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
        elif activation_fn == "geglu":
            act_fn = GEGLU(dim, inner_dim, bias=bias)
        elif activation_fn == "geglu-approximate":
            act_fn = ApproximateGELU(dim, inner_dim, bias=bias)

        self.net = nn.ModuleList([])
        # project in
        self.net.append(act_fn)
        # project dropout
        self.net.append(nn.Dropout(dropout))
        # project out
        self.net.append(linear_cls(inner_dim, dim_out, bias=bias))
        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
        if final_dropout:
            self.net.append(nn.Dropout(dropout))

    def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
        compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear)
        for module in self.net:
            if isinstance(module, compatible_cls):
                hidden_states = module(hidden_states, scale)
            else:
                hidden_states = module(hidden_states)
        return hidden_states