videosys/models/open_sora_plan/latte.py

# Adapted from Open-Sora-Plan

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# References:
# Open-Sora-Plan: https://github.com/PKU-YuanGroup/Open-Sora-Plan
# --------------------------------------------------------


import json
import os
from dataclasses import dataclass
from functools import partial
from importlib import import_module
from typing import Any, Callable, Dict, Optional, Tuple

import numpy as np
import torch
import torch.nn.functional as F
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.activations import GEGLU, GELU, ApproximateGELU
from diffusers.models.attention_processor import (
    AttnAddedKVProcessor,
    AttnAddedKVProcessor2_0,
    AttnProcessor,
    CustomDiffusionAttnProcessor,
    CustomDiffusionAttnProcessor2_0,
    CustomDiffusionXFormersAttnProcessor,
    LoRAAttnAddedKVProcessor,
    LoRAAttnProcessor,
    LoRAAttnProcessor2_0,
    LoRAXFormersAttnProcessor,
    SlicedAttnAddedKVProcessor,
    SlicedAttnProcessor,
    SpatialNorm,
    XFormersAttnAddedKVProcessor,
    XFormersAttnProcessor,
)
from diffusers.models.embeddings import SinusoidalPositionalEmbedding, TimestepEmbedding, Timesteps
from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormZero
from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_xformers_available
from diffusers.utils.torch_utils import maybe_allow_in_graph
from einops import rearrange, repeat
from torch import nn

from videosys.core.comm import (
    all_to_all_with_pad,
    gather_sequence,
    get_spatial_pad,
    get_temporal_pad,
    set_spatial_pad,
    set_temporal_pad,
    split_sequence,
)
from videosys.core.pab_mgr import (
    enable_pab,
    get_mlp_output,
    if_broadcast_cross,
    if_broadcast_mlp,
    if_broadcast_spatial,
    if_broadcast_temporal,
    save_mlp_output,
)
from videosys.core.parallel_mgr import (
    enable_sequence_parallel,
    get_cfg_parallel_group,
    get_cfg_parallel_size,
    get_sequence_parallel_group,
)
from videosys.utils.logging import logger
from videosys.utils.utils import batch_func

if is_xformers_available():
    import xformers
    import xformers.ops
else:
    xformers = None


SPATIAL_LIST = []
TEMPROAL_LIST = []
CROSS_LIST = []


def get_2d_sincos_pos_embed(
    embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16
):
    """
    grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or
    [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
    """
    if isinstance(grid_size, int):
        grid_size = (grid_size, grid_size)

    grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0] / base_size) / interpolation_scale
    grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1] / base_size) / interpolation_scale
    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
    grid = np.stack(grid, axis=0)

    grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])
    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
    if cls_token and extra_tokens > 0:
        pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
    return pos_embed


def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
    if embed_dim % 2 != 0:
        raise ValueError("embed_dim must be divisible by 2")

    # use half of dimensions to encode grid_h
    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)

    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
    return emb


def get_1d_sincos_pos_embed(embed_dim, length, interpolation_scale=1.0, base_size=16):
    pos = torch.arange(0, length).unsqueeze(1) / interpolation_scale
    pos_embed = get_1d_sincos_pos_embed_from_grid(embed_dim, pos)
    return pos_embed


def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    """
    embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)
    """
    if embed_dim % 2 != 0:
        raise ValueError("embed_dim must be divisible by 2")

    omega = np.arange(embed_dim // 2, dtype=np.float64)
    omega /= embed_dim / 2.0
    omega = 1.0 / 10000**omega  # (D/2,)

    pos = pos.reshape(-1)  # (M,)
    out = np.einsum("m,d->md", pos, omega)  # (M, D/2), outer product

    emb_sin = np.sin(out)  # (M, D/2)
    emb_cos = np.cos(out)  # (M, D/2)

    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
    return emb


class RoPE2D(torch.nn.Module):
    def __init__(self, freq=10000.0, F0=1.0, scaling_factor=1.0):
        super().__init__()
        self.base = freq
        self.F0 = F0
        self.scaling_factor = scaling_factor
        self.cache = {}

    def get_cos_sin(self, D, seq_len, device, dtype):
        if (D, seq_len, device, dtype) not in self.cache:
            inv_freq = 1.0 / (self.base ** (torch.arange(0, D, 2).float().to(device) / D))
            t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)
            freqs = torch.einsum("i,j->ij", t, inv_freq).to(dtype)
            freqs = torch.cat((freqs, freqs), dim=-1)
            cos = freqs.cos()  # (Seq, Dim)
            sin = freqs.sin()
            self.cache[D, seq_len, device, dtype] = (cos, sin)
        return self.cache[D, seq_len, device, dtype]

    @staticmethod
    def rotate_half(x):
        x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
        return torch.cat((-x2, x1), dim=-1)

    def apply_rope1d(self, tokens, pos1d, cos, sin):
        assert pos1d.ndim == 2

        cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :]
        sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :]
        return (tokens * cos) + (self.rotate_half(tokens) * sin)

    def forward(self, tokens, positions):
        """
        input:
            * tokens: batch_size x nheads x ntokens x dim
            * positions: batch_size x ntokens x 2 (y and x position of each token)
        output:
            * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim)
        """
        assert tokens.size(3) % 2 == 0, "number of dimensions should be a multiple of two"
        D = tokens.size(3) // 2
        assert positions.ndim == 3 and positions.shape[-1] == 2  # Batch, Seq, 2
        cos, sin = self.get_cos_sin(D, int(positions.max()) + 1, tokens.device, tokens.dtype)
        # split features into two along the feature dimension, and apply rope1d on each half
        y, x = tokens.chunk(2, dim=-1)
        y = self.apply_rope1d(y, positions[:, :, 0], cos, sin)
        x = self.apply_rope1d(x, positions[:, :, 1], cos, sin)
        tokens = torch.cat((y, x), dim=-1)
        return tokens


class LinearScalingRoPE2D(RoPE2D):
    """Code from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L148"""

    def forward(self, tokens, positions):
        # difference to the original RoPE: a scaling factor is aplied to the position ids
        dtype = positions.dtype
        positions = positions.float() / self.scaling_factor
        positions = positions.to(dtype)
        tokens = super().forward(tokens, positions)
        return tokens


class RoPE1D(torch.nn.Module):
    def __init__(self, freq=10000.0, F0=1.0, scaling_factor=1.0):
        super().__init__()
        self.base = freq
        self.F0 = F0
        self.scaling_factor = scaling_factor
        self.cache = {}

    def get_cos_sin(self, D, seq_len, device, dtype):
        if (D, seq_len, device, dtype) not in self.cache:
            inv_freq = 1.0 / (self.base ** (torch.arange(0, D, 2).float().to(device) / D))
            t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)
            freqs = torch.einsum("i,j->ij", t, inv_freq).to(dtype)
            freqs = torch.cat((freqs, freqs), dim=-1)
            cos = freqs.cos()  # (Seq, Dim)
            sin = freqs.sin()
            self.cache[D, seq_len, device, dtype] = (cos, sin)
        return self.cache[D, seq_len, device, dtype]

    @staticmethod
    def rotate_half(x):
        x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
        return torch.cat((-x2, x1), dim=-1)

    def apply_rope1d(self, tokens, pos1d, cos, sin):
        assert pos1d.ndim == 2
        cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :]
        sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :]
        return (tokens * cos) + (self.rotate_half(tokens) * sin)

    def forward(self, tokens, positions):
        """
        input:
            * tokens: batch_size x nheads x ntokens x dim
            * positions: batch_size x ntokens (t position of each token)
        output:
            * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim)
        """
        D = tokens.size(3)
        assert positions.ndim == 2  # Batch, Seq
        cos, sin = self.get_cos_sin(D, int(positions.max()) + 1, tokens.device, tokens.dtype)
        tokens = self.apply_rope1d(tokens, positions, cos, sin)
        return tokens


class LinearScalingRoPE1D(RoPE1D):
    """Code from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L148"""

    def forward(self, tokens, positions):
        # difference to the original RoPE: a scaling factor is aplied to the position ids
        dtype = positions.dtype
        positions = positions.float() / self.scaling_factor
        positions = positions.to(dtype)
        tokens = super().forward(tokens, positions)
        return tokens


class PositionGetter2D(object):
    """return positions of patches"""

    def __init__(self):
        self.cache_positions = {}

    def __call__(self, b, h, w, device):
        if not (h, w) in self.cache_positions:
            x = torch.arange(w, device=device)
            y = torch.arange(h, device=device)
            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)
        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()
        return pos


class PositionGetter1D(object):
    """return positions of patches"""

    def __init__(self):
        self.cache_positions = {}

    def __call__(self, b, l, device):
        if not (l) in self.cache_positions:
            x = torch.arange(l, device=device)
            self.cache_positions[l] = x  # (l, )
        pos = self.cache_positions[l].view(1, l).expand(b, -1).clone()
        return pos


class CombinedTimestepSizeEmbeddings(nn.Module):
    """
    For PixArt-Alpha.

    Reference:
    https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L164C9-L168C29
    """

    def __init__(self, embedding_dim, size_emb_dim, use_additional_conditions: bool = False):
        super().__init__()

        self.outdim = size_emb_dim
        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)

        self.use_additional_conditions = use_additional_conditions
        if use_additional_conditions:
            self.use_additional_conditions = True
            self.additional_condition_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
            self.resolution_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=size_emb_dim)
            self.aspect_ratio_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=size_emb_dim)

    def apply_condition(self, size: torch.Tensor, batch_size: int, embedder: nn.Module):
        if size.ndim == 1:
            size = size[:, None]

        if size.shape[0] != batch_size:
            size = size.repeat(batch_size // size.shape[0], 1)
            if size.shape[0] != batch_size:
                raise ValueError(f"`batch_size` should be {size.shape[0]} but found {batch_size}.")

        current_batch_size, dims = size.shape[0], size.shape[1]
        size = size.reshape(-1)
        size_freq = self.additional_condition_proj(size).to(size.dtype)

        size_emb = embedder(size_freq)
        size_emb = size_emb.reshape(current_batch_size, dims * self.outdim)
        return size_emb

    def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype):
        timesteps_proj = self.time_proj(timestep)
        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)

        if self.use_additional_conditions:
            resolution = self.apply_condition(resolution, batch_size=batch_size, embedder=self.resolution_embedder)
            aspect_ratio = self.apply_condition(
                aspect_ratio, batch_size=batch_size, embedder=self.aspect_ratio_embedder
            )
            conditioning = timesteps_emb + torch.cat([resolution, aspect_ratio], dim=1)
        else:
            conditioning = timesteps_emb

        return conditioning


class CaptionProjection(nn.Module):
    """
    Projects caption embeddings. Also handles dropout for classifier-free guidance.

    Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
    """

    def __init__(self, in_features, hidden_size, num_tokens=120):
        super().__init__()
        self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True)
        self.act_1 = nn.GELU(approximate="tanh")
        self.linear_2 = nn.Linear(in_features=hidden_size, out_features=hidden_size, bias=True)
        self.register_buffer("y_embedding", nn.Parameter(torch.randn(num_tokens, in_features) / in_features**0.5))

    def forward(self, caption, force_drop_ids=None):
        hidden_states = self.linear_1(caption)
        hidden_states = self.act_1(hidden_states)
        hidden_states = self.linear_2(hidden_states)
        return hidden_states


class PatchEmbed(nn.Module):
    """2D Image to Patch Embedding"""

    def __init__(
        self,
        height=224,
        width=224,
        patch_size=16,
        in_channels=3,
        embed_dim=768,
        layer_norm=False,
        flatten=True,
        bias=True,
        interpolation_scale=1,
    ):
        super().__init__()

        num_patches = (height // patch_size) * (width // patch_size)
        self.flatten = flatten
        self.layer_norm = layer_norm

        self.proj = nn.Conv2d(
            in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
        )
        if layer_norm:
            self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6)
        else:
            self.norm = None

        self.patch_size = patch_size
        # See:
        # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L161
        self.height, self.width = height // patch_size, width // patch_size

        self.base_size = height // patch_size
        self.interpolation_scale = interpolation_scale
        pos_embed = get_2d_sincos_pos_embed(
            embed_dim, int(num_patches**0.5), base_size=self.base_size, interpolation_scale=self.interpolation_scale
        )
        self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)

    def forward(self, latent):
        height, width = latent.shape[-2] // self.patch_size, latent.shape[-1] // self.patch_size

        latent = self.proj(latent)
        if self.flatten:
            latent = latent.flatten(2).transpose(1, 2)  # BCHW -> BNC
        if self.layer_norm:
            latent = self.norm(latent)
        # Interpolate positional embeddings if needed.
        # (For PixArt-Alpha: https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L162C151-L162C160)
        if self.height != height or self.width != width:
            # raise ValueError
            pos_embed = get_2d_sincos_pos_embed(
                embed_dim=self.pos_embed.shape[-1],
                grid_size=(height, width),
                base_size=self.base_size,
                interpolation_scale=self.interpolation_scale,
            )
            pos_embed = torch.from_numpy(pos_embed)
            pos_embed = pos_embed.float().unsqueeze(0).to(latent.device)
        else:
            pos_embed = self.pos_embed
        return (latent + pos_embed).to(latent.dtype)


@maybe_allow_in_graph
class Attention(nn.Module):
    r"""
    A cross attention layer.

    Parameters:
        query_dim (`int`):
            The number of channels in the query.
        cross_attention_dim (`int`, *optional*):
            The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
        heads (`int`,  *optional*, defaults to 8):
            The number of heads to use for multi-head attention.
        dim_head (`int`,  *optional*, defaults to 64):
            The number of channels in each head.
        dropout (`float`, *optional*, defaults to 0.0):
            The dropout probability to use.
        bias (`bool`, *optional*, defaults to False):
            Set to `True` for the query, key, and value linear layers to contain a bias parameter.
        upcast_attention (`bool`, *optional*, defaults to False):
            Set to `True` to upcast the attention computation to `float32`.
        upcast_softmax (`bool`, *optional*, defaults to False):
            Set to `True` to upcast the softmax computation to `float32`.
        cross_attention_norm (`str`, *optional*, defaults to `None`):
            The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.
        cross_attention_norm_num_groups (`int`, *optional*, defaults to 32):
            The number of groups to use for the group norm in the cross attention.
        added_kv_proj_dim (`int`, *optional*, defaults to `None`):
            The number of channels to use for the added key and value projections. If `None`, no projection is used.
        norm_num_groups (`int`, *optional*, defaults to `None`):
            The number of groups to use for the group norm in the attention.
        spatial_norm_dim (`int`, *optional*, defaults to `None`):
            The number of channels to use for the spatial normalization.
        out_bias (`bool`, *optional*, defaults to `True`):
            Set to `True` to use a bias in the output linear layer.
        scale_qk (`bool`, *optional*, defaults to `True`):
            Set to `True` to scale the query and key by `1 / sqrt(dim_head)`.
        only_cross_attention (`bool`, *optional*, defaults to `False`):
            Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if
            `added_kv_proj_dim` is not `None`.
        eps (`float`, *optional*, defaults to 1e-5):
            An additional value added to the denominator in group normalization that is used for numerical stability.
        rescale_output_factor (`float`, *optional*, defaults to 1.0):
            A factor to rescale the output by dividing it with this value.
        residual_connection (`bool`, *optional*, defaults to `False`):
            Set to `True` to add the residual connection to the output.
        _from_deprecated_attn_block (`bool`, *optional*, defaults to `False`):
            Set to `True` if the attention block is loaded from a deprecated state dict.
        processor (`AttnProcessor`, *optional*, defaults to `None`):
            The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and
            `AttnProcessor` otherwise.
    """

    def __init__(
        self,
        query_dim: int,
        cross_attention_dim: Optional[int] = None,
        heads: int = 8,
        dim_head: int = 64,
        dropout: float = 0.0,
        bias: bool = False,
        upcast_attention: bool = False,
        upcast_softmax: bool = False,
        cross_attention_norm: Optional[str] = None,
        cross_attention_norm_num_groups: int = 32,
        added_kv_proj_dim: Optional[int] = None,
        norm_num_groups: Optional[int] = None,
        spatial_norm_dim: Optional[int] = None,
        out_bias: bool = True,
        scale_qk: bool = True,
        only_cross_attention: bool = False,
        eps: float = 1e-5,
        rescale_output_factor: float = 1.0,
        residual_connection: bool = False,
        _from_deprecated_attn_block: bool = False,
        processor: Optional["AttnProcessor"] = None,
        attention_mode: str = "xformers",
        use_rope: bool = False,
        rope_scaling: Optional[Dict] = None,
        compress_kv_factor: Optional[Tuple] = None,
    ):
        super().__init__()
        self.inner_dim = dim_head * heads
        self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
        self.upcast_attention = upcast_attention
        self.upcast_softmax = upcast_softmax
        self.rescale_output_factor = rescale_output_factor
        self.residual_connection = residual_connection
        self.dropout = dropout
        self.use_rope = use_rope
        self.rope_scaling = rope_scaling
        self.compress_kv_factor = compress_kv_factor

        # we make use of this private variable to know whether this class is loaded
        # with an deprecated state dict so that we can convert it on the fly
        self._from_deprecated_attn_block = _from_deprecated_attn_block

        self.scale_qk = scale_qk
        self.scale = dim_head**-0.5 if self.scale_qk else 1.0

        self.heads = heads
        # for slice_size > 0 the attention score computation
        # is split across the batch axis to save memory
        # You can set slice_size with `set_attention_slice`
        self.sliceable_head_dim = heads

        self.added_kv_proj_dim = added_kv_proj_dim
        self.only_cross_attention = only_cross_attention

        if self.added_kv_proj_dim is None and self.only_cross_attention:
            raise ValueError(
                "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."
            )

        if norm_num_groups is not None:
            self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True)
        else:
            self.group_norm = None

        if spatial_norm_dim is not None:
            self.spatial_norm = SpatialNorm(f_channels=query_dim, zq_channels=spatial_norm_dim)
        else:
            self.spatial_norm = None

        if cross_attention_norm is None:
            self.norm_cross = None
        elif cross_attention_norm == "layer_norm":
            self.norm_cross = nn.LayerNorm(self.cross_attention_dim)
        elif cross_attention_norm == "group_norm":
            if self.added_kv_proj_dim is not None:
                # The given `encoder_hidden_states` are initially of shape
                # (batch_size, seq_len, added_kv_proj_dim) before being projected
                # to (batch_size, seq_len, cross_attention_dim). The norm is applied
                # before the projection, so we need to use `added_kv_proj_dim` as
                # the number of channels for the group norm.
                norm_cross_num_channels = added_kv_proj_dim
            else:
                norm_cross_num_channels = self.cross_attention_dim

            self.norm_cross = nn.GroupNorm(
                num_channels=norm_cross_num_channels, num_groups=cross_attention_norm_num_groups, eps=1e-5, affine=True
            )
        else:
            raise ValueError(
                f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'"
            )

        if USE_PEFT_BACKEND:
            linear_cls = nn.Linear
        else:
            linear_cls = LoRACompatibleLinear

        assert not (
            self.use_rope and (self.compress_kv_factor is not None)
        ), "Can not both enable compressing kv and using rope"
        if self.compress_kv_factor is not None:
            self._init_compress()

        self.to_q = linear_cls(query_dim, self.inner_dim, bias=bias)

        if not self.only_cross_attention:
            # only relevant for the `AddedKVProcessor` classes
            self.to_k = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias)
            self.to_v = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias)
        else:
            self.to_k = None
            self.to_v = None

        if self.added_kv_proj_dim is not None:
            self.add_k_proj = linear_cls(added_kv_proj_dim, self.inner_dim)
            self.add_v_proj = linear_cls(added_kv_proj_dim, self.inner_dim)

        self.to_out = nn.ModuleList([])
        self.to_out.append(linear_cls(self.inner_dim, query_dim, bias=out_bias))
        self.to_out.append(nn.Dropout(dropout))

        # set attention processor
        # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
        # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
        # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
        if processor is None:
            processor = (
                AttnProcessor2_0(
                    self.inner_dim,
                    attention_mode,
                    use_rope,
                    rope_scaling=rope_scaling,
                    compress_kv_factor=compress_kv_factor,
                )
                if hasattr(F, "scaled_dot_product_attention") and self.scale_qk
                else AttnProcessor()
            )
        self.set_processor(processor)

    def set_use_memory_efficient_attention_xformers(
        self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None
    ) -> None:
        r"""
        Set whether to use memory efficient attention from `xformers` or not.

        Args:
            use_memory_efficient_attention_xformers (`bool`):
                Whether to use memory efficient attention from `xformers` or not.
            attention_op (`Callable`, *optional*):
                The attention operation to use. Defaults to `None` which uses the default attention operation from
                `xformers`.
        """
        is_lora = hasattr(self, "processor")
        is_custom_diffusion = hasattr(self, "processor") and isinstance(
            self.processor,
            (CustomDiffusionAttnProcessor, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0),
        )
        is_added_kv_processor = hasattr(self, "processor") and isinstance(
            self.processor,
            (
                AttnAddedKVProcessor,
                AttnAddedKVProcessor2_0,
                SlicedAttnAddedKVProcessor,
                XFormersAttnAddedKVProcessor,
                LoRAAttnAddedKVProcessor,
            ),
        )

        if use_memory_efficient_attention_xformers:
            if is_added_kv_processor and (is_lora or is_custom_diffusion):
                raise NotImplementedError(
                    f"Memory efficient attention is currently not supported for LoRA or custom diffusion for attention processor type {self.processor}"
                )
            if not is_xformers_available():
                raise ModuleNotFoundError(
                    (
                        "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
                        " xformers"
                    ),
                    name="xformers",
                )
            elif not torch.cuda.is_available():
                raise ValueError(
                    "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
                    " only available for GPU "
                )
            else:
                try:
                    # Make sure we can run the memory efficient attention
                    _ = xformers.ops.memory_efficient_attention(
                        torch.randn((1, 2, 40), device="cuda"),
                        torch.randn((1, 2, 40), device="cuda"),
                        torch.randn((1, 2, 40), device="cuda"),
                    )
                except Exception as e:
                    raise e

            if is_lora:
                # TODO (sayakpaul): should we throw a warning if someone wants to use the xformers
                # variant when using PT 2.0 now that we have LoRAAttnProcessor2_0?
                processor = LoRAXFormersAttnProcessor(
                    hidden_size=self.processor.hidden_size,
                    cross_attention_dim=self.processor.cross_attention_dim,
                    rank=self.processor.rank,
                    attention_op=attention_op,
                )
                processor.load_state_dict(self.processor.state_dict())
                processor.to(self.processor.to_q_lora.up.weight.device)
            elif is_custom_diffusion:
                processor = CustomDiffusionXFormersAttnProcessor(
                    train_kv=self.processor.train_kv,
                    train_q_out=self.processor.train_q_out,
                    hidden_size=self.processor.hidden_size,
                    cross_attention_dim=self.processor.cross_attention_dim,
                    attention_op=attention_op,
                )
                processor.load_state_dict(self.processor.state_dict())
                if hasattr(self.processor, "to_k_custom_diffusion"):
                    processor.to(self.processor.to_k_custom_diffusion.weight.device)
            elif is_added_kv_processor:
                # TODO(Patrick, Suraj, William) - currently xformers doesn't work for UnCLIP
                # which uses this type of cross attention ONLY because the attention mask of format
                # [0, ..., -10.000, ..., 0, ...,] is not supported
                # throw warning
                logger.info(
                    "Memory efficient attention with `xformers` might currently not work correctly if an attention mask is required for the attention operation."
                )
                processor = XFormersAttnAddedKVProcessor(attention_op=attention_op)
            else:
                processor = XFormersAttnProcessor(attention_op=attention_op)
        else:
            if is_lora:
                attn_processor_class = (
                    LoRAAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else LoRAAttnProcessor
                )
                processor = attn_processor_class(
                    hidden_size=self.processor.hidden_size,
                    cross_attention_dim=self.processor.cross_attention_dim,
                    rank=self.processor.rank,
                )
                processor.load_state_dict(self.processor.state_dict())
                processor.to(self.processor.to_q_lora.up.weight.device)
            elif is_custom_diffusion:
                attn_processor_class = (
                    CustomDiffusionAttnProcessor2_0
                    if hasattr(F, "scaled_dot_product_attention")
                    else CustomDiffusionAttnProcessor
                )
                processor = attn_processor_class(
                    train_kv=self.processor.train_kv,
                    train_q_out=self.processor.train_q_out,
                    hidden_size=self.processor.hidden_size,
                    cross_attention_dim=self.processor.cross_attention_dim,
                )
                processor.load_state_dict(self.processor.state_dict())
                if hasattr(self.processor, "to_k_custom_diffusion"):
                    processor.to(self.processor.to_k_custom_diffusion.weight.device)
            else:
                # set attention processor
                # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
                # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
                # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
                processor = (
                    AttnProcessor2_0()
                    if hasattr(F, "scaled_dot_product_attention") and self.scale_qk
                    else AttnProcessor()
                )

        self.set_processor(processor)

    def set_attention_slice(self, slice_size: int) -> None:
        r"""
        Set the slice size for attention computation.

        Args:
            slice_size (`int`):
                The slice size for attention computation.
        """
        if slice_size is not None and slice_size > self.sliceable_head_dim:
            raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")

        if slice_size is not None and self.added_kv_proj_dim is not None:
            processor = SlicedAttnAddedKVProcessor(slice_size)
        elif slice_size is not None:
            processor = SlicedAttnProcessor(slice_size)
        elif self.added_kv_proj_dim is not None:
            processor = AttnAddedKVProcessor()
        else:
            # set attention processor
            # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
            # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
            # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
            processor = (
                AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
            )

        self.set_processor(processor)

    def set_processor(self, processor: "AttnProcessor", _remove_lora: bool = False) -> None:
        r"""
        Set the attention processor to use.

        Args:
            processor (`AttnProcessor`):
                The attention processor to use.
            _remove_lora (`bool`, *optional*, defaults to `False`):
                Set to `True` to remove LoRA layers from the model.
        """
        if not USE_PEFT_BACKEND and hasattr(self, "processor") and _remove_lora and self.to_q.lora_layer is not None:
            deprecate(
                "set_processor to offload LoRA",
                "0.26.0",
                "In detail, removing LoRA layers via calling `set_default_attn_processor` is deprecated. Please make sure to call `pipe.unload_lora_weights()` instead.",
            )
            # TODO(Patrick, Sayak) - this can be deprecated once PEFT LoRA integration is complete
            # We need to remove all LoRA layers
            # Don't forget to remove ALL `_remove_lora` from the codebase
            for module in self.modules():
                if hasattr(module, "set_lora_layer"):
                    module.set_lora_layer(None)

        # if current processor is in `self._modules` and if passed `processor` is not, we need to
        # pop `processor` from `self._modules`
        if (
            hasattr(self, "processor")
            and isinstance(self.processor, torch.nn.Module)
            and not isinstance(processor, torch.nn.Module)
        ):
            logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}")
            self._modules.pop("processor")

        self.processor = processor

    def get_processor(self, return_deprecated_lora: bool = False):
        r"""
        Get the attention processor in use.

        Args:
            return_deprecated_lora (`bool`, *optional*, defaults to `False`):
                Set to `True` to return the deprecated LoRA attention processor.

        Returns:
            "AttentionProcessor": The attention processor in use.
        """
        if not return_deprecated_lora:
            return self.processor

        # TODO(Sayak, Patrick). The rest of the function is needed to ensure backwards compatible
        # serialization format for LoRA Attention Processors. It should be deleted once the integration
        # with PEFT is completed.
        is_lora_activated = {
            name: module.lora_layer is not None
            for name, module in self.named_modules()
            if hasattr(module, "lora_layer")
        }

        # 1. if no layer has a LoRA activated we can return the processor as usual
        if not any(is_lora_activated.values()):
            return self.processor

        # If doesn't apply LoRA do `add_k_proj` or `add_v_proj`
        is_lora_activated.pop("add_k_proj", None)
        is_lora_activated.pop("add_v_proj", None)
        # 2. else it is not posssible that only some layers have LoRA activated
        if not all(is_lora_activated.values()):
            raise ValueError(
                f"Make sure that either all layers or no layers have LoRA activated, but have {is_lora_activated}"
            )

        # 3. And we need to merge the current LoRA layers into the corresponding LoRA attention processor
        non_lora_processor_cls_name = self.processor.__class__.__name__
        lora_processor_cls = getattr(import_module(__name__), "LoRA" + non_lora_processor_cls_name)

        hidden_size = self.inner_dim

        # now create a LoRA attention processor from the LoRA layers
        if lora_processor_cls in [LoRAAttnProcessor, LoRAAttnProcessor2_0, LoRAXFormersAttnProcessor]:
            kwargs = {
                "cross_attention_dim": self.cross_attention_dim,
                "rank": self.to_q.lora_layer.rank,
                "network_alpha": self.to_q.lora_layer.network_alpha,
                "q_rank": self.to_q.lora_layer.rank,
                "q_hidden_size": self.to_q.lora_layer.out_features,
                "k_rank": self.to_k.lora_layer.rank,
                "k_hidden_size": self.to_k.lora_layer.out_features,
                "v_rank": self.to_v.lora_layer.rank,
                "v_hidden_size": self.to_v.lora_layer.out_features,
                "out_rank": self.to_out[0].lora_layer.rank,
                "out_hidden_size": self.to_out[0].lora_layer.out_features,
            }

            if hasattr(self.processor, "attention_op"):
                kwargs["attention_op"] = self.processor.attention_op

            lora_processor = lora_processor_cls(hidden_size, **kwargs)
            lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict())
            lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict())
            lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict())
            lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict())
        elif lora_processor_cls == LoRAAttnAddedKVProcessor:
            lora_processor = lora_processor_cls(
                hidden_size,
                cross_attention_dim=self.add_k_proj.weight.shape[0],
                rank=self.to_q.lora_layer.rank,
                network_alpha=self.to_q.lora_layer.network_alpha,
            )
            lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict())
            lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict())
            lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict())
            lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict())

            # only save if used
            if self.add_k_proj.lora_layer is not None:
                lora_processor.add_k_proj_lora.load_state_dict(self.add_k_proj.lora_layer.state_dict())
                lora_processor.add_v_proj_lora.load_state_dict(self.add_v_proj.lora_layer.state_dict())
            else:
                lora_processor.add_k_proj_lora = None
                lora_processor.add_v_proj_lora = None
        else:
            raise ValueError(f"{lora_processor_cls} does not exist.")

        return lora_processor

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        **cross_attention_kwargs,
    ) -> torch.Tensor:
        r"""
        The forward method of the `Attention` class.

        Args:
            hidden_states (`torch.Tensor`):
                The hidden states of the query.
            encoder_hidden_states (`torch.Tensor`, *optional*):
                The hidden states of the encoder.
            attention_mask (`torch.Tensor`, *optional*):
                The attention mask to use. If `None`, no mask is applied.
            **cross_attention_kwargs:
                Additional keyword arguments to pass along to the cross attention.

        Returns:
            `torch.Tensor`: The output of the attention layer.
        """
        # The `Attention` class can call different attention processors / attention functions
        # here we simply pass along all tensors to the selected processor class
        # For standard processors that are defined here, `**cross_attention_kwargs` is empty
        return self.processor(
            self,
            hidden_states,
            encoder_hidden_states=encoder_hidden_states,
            attention_mask=attention_mask,
            **cross_attention_kwargs,
        )

    def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor:
        r"""
        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`. `heads`
        is the number of heads initialized while constructing the `Attention` class.

        Args:
            tensor (`torch.Tensor`): The tensor to reshape.

        Returns:
            `torch.Tensor`: The reshaped tensor.
        """
        head_size = self.heads
        batch_size, seq_len, dim = tensor.shape
        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
        return tensor

    def head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor:
        r"""
        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size, seq_len, heads, dim // heads]` `heads` is
        the number of heads initialized while constructing the `Attention` class.

        Args:
            tensor (`torch.Tensor`): The tensor to reshape.
            out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor. If `3`, the tensor is
                reshaped to `[batch_size * heads, seq_len, dim // heads]`.

        Returns:
            `torch.Tensor`: The reshaped tensor.
        """
        head_size = self.heads
        batch_size, seq_len, dim = tensor.shape
        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
        tensor = tensor.permute(0, 2, 1, 3)

        if out_dim == 3:
            tensor = tensor.reshape(batch_size * head_size, seq_len, dim // head_size)

        return tensor

    def get_attention_scores(
        self, query: torch.Tensor, key: torch.Tensor, attention_mask: torch.Tensor = None
    ) -> torch.Tensor:
        r"""
        Compute the attention scores.

        Args:
            query (`torch.Tensor`): The query tensor.
            key (`torch.Tensor`): The key tensor.
            attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied.

        Returns:
            `torch.Tensor`: The attention probabilities/scores.
        """
        dtype = query.dtype
        if self.upcast_attention:
            query = query.float()
            key = key.float()

        if attention_mask is None:
            baddbmm_input = torch.empty(
                query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device
            )
            beta = 0
        else:
            baddbmm_input = attention_mask
            beta = 1

        attention_scores = torch.baddbmm(
            baddbmm_input,
            query,
            key.transpose(-1, -2),
            beta=beta,
            alpha=self.scale,
        )
        del baddbmm_input

        if self.upcast_softmax:
            attention_scores = attention_scores.float()

        attention_probs = attention_scores.softmax(dim=-1)
        del attention_scores

        attention_probs = attention_probs.to(dtype)

        return attention_probs

    def prepare_attention_mask(
        self, attention_mask: torch.Tensor, target_length: int, batch_size: int, out_dim: int = 3
    ) -> torch.Tensor:
        r"""
        Prepare the attention mask for the attention computation.

        Args:
            attention_mask (`torch.Tensor`):
                The attention mask to prepare.
            target_length (`int`):
                The target length of the attention mask. This is the length of the attention mask after padding.
            batch_size (`int`):
                The batch size, which is used to repeat the attention mask.
            out_dim (`int`, *optional*, defaults to `3`):
                The output dimension of the attention mask. Can be either `3` or `4`.

        Returns:
            `torch.Tensor`: The prepared attention mask.
        """
        head_size = self.heads
        if attention_mask is None:
            return attention_mask

        current_length: int = attention_mask.shape[-1]
        if current_length != target_length:
            if attention_mask.device.type == "mps":
                # HACK: MPS: Does not support padding by greater than dimension of input tensor.
                # Instead, we can manually construct the padding tensor.
                padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length)
                padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device)
                attention_mask = torch.cat([attention_mask, padding], dim=2)
            else:
                # TODO: for pipelines such as stable-diffusion, padding cross-attn mask:
                #       we want to instead pad by (0, remaining_length), where remaining_length is:
                #       remaining_length: int = target_length - current_length
                # TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding
                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)

        if out_dim == 3:
            if attention_mask.shape[0] < batch_size * head_size:
                attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
        elif out_dim == 4:
            attention_mask = attention_mask.unsqueeze(1)
            attention_mask = attention_mask.repeat_interleave(head_size, dim=1)

        return attention_mask

    def norm_encoder_hidden_states(self, encoder_hidden_states: torch.Tensor) -> torch.Tensor:
        r"""
        Normalize the encoder hidden states. Requires `self.norm_cross` to be specified when constructing the
        `Attention` class.

        Args:
            encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder.

        Returns:
            `torch.Tensor`: The normalized encoder hidden states.
        """
        assert self.norm_cross is not None, "self.norm_cross must be defined to call self.norm_encoder_hidden_states"

        if isinstance(self.norm_cross, nn.LayerNorm):
            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
        elif isinstance(self.norm_cross, nn.GroupNorm):
            # Group norm norms along the channels dimension and expects
            # input to be in the shape of (N, C, *). In this case, we want
            # to norm along the hidden dimension, so we need to move
            # (batch_size, sequence_length, hidden_size) ->
            # (batch_size, hidden_size, sequence_length)
            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
        else:
            assert False

        return encoder_hidden_states

    def _init_compress(self):
        if len(self.compress_kv_factor) == 2:
            self.sr = nn.Conv2d(
                self.inner_dim,
                self.inner_dim,
                groups=self.inner_dim,
                kernel_size=self.compress_kv_factor,
                stride=self.compress_kv_factor,
            )
            self.sr.weight.data.fill_(1 / self.compress_kv_factor[0] ** 2)
        elif len(self.compress_kv_factor) == 1:
            self.kernel_size = self.compress_kv_factor[0]
            self.sr = nn.Conv1d(
                self.inner_dim,
                self.inner_dim,
                groups=self.inner_dim,
                kernel_size=self.compress_kv_factor[0],
                stride=self.compress_kv_factor[0],
            )
            self.sr.weight.data.fill_(1 / self.compress_kv_factor[0])
        self.sr.bias.data.zero_()
        self.norm = nn.LayerNorm(self.inner_dim)


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

    def __init__(self, dim=1152, attention_mode="xformers", use_rope=False, rope_scaling=None, compress_kv_factor=None):
        self.dim = dim
        self.attention_mode = attention_mode
        self.use_rope = use_rope
        self.rope_scaling = rope_scaling
        self.compress_kv_factor = compress_kv_factor
        if self.use_rope:
            self._init_rope()

        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 _init_rope(self):
        if self.rope_scaling is None:
            self.rope2d = RoPE2D()
            self.rope1d = RoPE1D()
        else:
            scaling_type = self.rope_scaling["type"]
            scaling_factor_2d = self.rope_scaling["factor_2d"]
            scaling_factor_1d = self.rope_scaling["factor_1d"]
            if scaling_type == "linear":
                self.rope2d = LinearScalingRoPE2D(scaling_factor=scaling_factor_2d)
                self.rope1d = LinearScalingRoPE1D(scaling_factor=scaling_factor_1d)
            else:
                raise ValueError(f"Unknown RoPE scaling type {scaling_type}")

    def __call__(
        self,
        attn: Attention,
        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,
        position_q: Optional[torch.LongTensor] = None,
        position_k: Optional[torch.LongTensor] = None,
        last_shape: Tuple[int] = None,
    ) -> torch.FloatTensor:
        residual = hidden_states

        args = () if USE_PEFT_BACKEND else (scale,)

        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)

        if self.compress_kv_factor is not None:
            batch_size = hidden_states.shape[0]
            if len(last_shape) == 2:
                encoder_hidden_states = hidden_states.permute(0, 2, 1).reshape(batch_size, self.dim, *last_shape)
                encoder_hidden_states = (
                    attn.sr(encoder_hidden_states).reshape(batch_size, self.dim, -1).permute(0, 2, 1)
                )
            elif len(last_shape) == 1:
                encoder_hidden_states = hidden_states.permute(0, 2, 1)
                if last_shape[0] % 2 == 1:
                    first_frame_pad = encoder_hidden_states[:, :, :1].repeat((1, 1, attn.kernel_size - 1))
                    encoder_hidden_states = torch.concatenate((first_frame_pad, encoder_hidden_states), dim=2)
                encoder_hidden_states = attn.sr(encoder_hidden_states).permute(0, 2, 1)
            else:
                raise NotImplementedError(f"NotImplementedError with last_shape {last_shape}")

            encoder_hidden_states = attn.norm(encoder_hidden_states)

        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)

        if self.use_rope:
            # require the shape of (batch_size x nheads x ntokens x dim)
            if position_q.ndim == 3:
                query = self.rope2d(query, position_q)
            elif position_q.ndim == 2:
                query = self.rope1d(query, position_q)
            else:
                raise NotImplementedError
            if position_k.ndim == 3:
                key = self.rope2d(key, position_k)
            elif position_k.ndim == 2:
                key = self.rope1d(key, position_k)
            else:
                raise NotImplementedError

        # the output of sdp = (batch, num_heads, seq_len, head_dim)
        # TODO: add support for attn.scale when we move to Torch 2.1
        if self.attention_mode == "flash":
            assert attention_mask is None or torch.all(
                attention_mask.bool()
            ), "flash-attn do not support attention_mask"
            with torch.backends.cuda.sdp_kernel(enable_math=False, enable_flash=True, enable_mem_efficient=False):
                hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
        elif self.attention_mode == "xformers":
            with torch.backends.cuda.sdp_kernel(enable_math=False, enable_flash=False, enable_mem_efficient=True):
                hidden_states = F.scaled_dot_product_attention(
                    query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
                )
        elif self.attention_mode == "math":
            hidden_states = F.scaled_dot_product_attention(
                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
            )
        else:
            raise NotImplementedError(f"Found attention_mode: {self.attention_mode}")
        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


@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


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.
    """

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

        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))
        # 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


@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,
        attention_mode: str = "xformers",
        use_rope: bool = False,
        rope_scaling: Optional[Dict] = None,
        compress_kv_factor: Optional[Tuple] = None,
        block_idx: Optional[int] = None,
    ):
        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"

        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)
        else:
            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)

        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 if only_cross_attention else None,
            upcast_attention=upcast_attention,
            attention_mode=attention_mode,
            use_rope=use_rope,
            rope_scaling=rope_scaling,
            compress_kv_factor=compress_kv_factor,
        )

        # # 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.
        #     self.norm2 = (
        #         AdaLayerNorm(dim, num_embeds_ada_norm)
        #         if self.use_ada_layer_norm
        #         else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
        #     )
        #     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,
        #     )  # is self-attn if encoder_hidden_states is none
        # else:
        #     self.norm2 = None
        #     self.attn2 = None

        # 3. Feed-forward
        # if not self.use_ada_layer_norm_single:
        #     self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)

        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)

        # 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

        # pab
        self.last_out = None
        self.count = 0
        self.block_idx = block_idx
        self.temp_mlp_count = 0

    def set_last_out(self, last_out: torch.Tensor):
        self.last_out = last_out

    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
        # 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,
        position_q: Optional[torch.LongTensor] = None,
        position_k: Optional[torch.LongTensor] = None,
        frame: int = None,
        org_timestep: Optional[torch.LongTensor] = None,
        all_timesteps=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]
        # 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)

        broadcast_temporal, self.count = if_broadcast_temporal(int(org_timestep[0]), self.count)
        if broadcast_temporal:
            attn_output = self.last_out
            assert 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)
        else:
            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_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)

            if enable_sequence_parallel():
                norm_hidden_states = self.dynamic_switch(norm_hidden_states, to_spatial_shard=True)

            attn_output = self.attn1(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
                attention_mask=attention_mask,
                position_q=position_q,
                position_k=position_k,
                last_shape=frame,
                **cross_attention_kwargs,
            )

            if enable_sequence_parallel():
                attn_output = self.dynamic_switch(attn_output, to_spatial_shard=False)

            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

            if enable_pab():
                self.set_last_out(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
        #     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 not self.use_ada_layer_norm_single:
        #     norm_hidden_states = self.norm3(hidden_states)

        if enable_pab():
            broadcast_mlp, self.temp_mlp_count, broadcast_next, broadcast_range = if_broadcast_mlp(
                int(org_timestep[0]),
                self.temp_mlp_count,
                self.block_idx,
                all_timesteps.tolist(),
                is_temporal=True,
            )

        if enable_pab() and broadcast_mlp:
            ff_output = get_mlp_output(
                broadcast_range,
                timestep=int(org_timestep[0]),
                block_idx=self.block_idx,
                is_temporal=True,
            )
        else:
            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 = self.norm3(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
                if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
                    raise ValueError(
                        f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
                    )

                num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
                ff_output = torch.cat(
                    [
                        self.ff(hid_slice, scale=lora_scale)
                        for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
                    ],
                    dim=self._chunk_dim,
                )
            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

            if enable_pab() and broadcast_next:
                save_mlp_output(
                    timestep=int(org_timestep[0]),
                    block_idx=self.block_idx,
                    ff_output=ff_output,
                    is_temporal=True,
                )

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

        return hidden_states

    def dynamic_switch(self, x, to_spatial_shard: bool):
        if to_spatial_shard:
            scatter_dim, gather_dim = 0, 1
            scatter_pad = get_spatial_pad()
            gather_pad = get_temporal_pad()
        else:
            scatter_dim, gather_dim = 1, 0
            scatter_pad = get_temporal_pad()
            gather_pad = get_spatial_pad()
        x = all_to_all_with_pad(
            x,
            get_sequence_parallel_group(),
            scatter_dim=scatter_dim,
            gather_dim=gather_dim,
            scatter_pad=scatter_pad,
            gather_pad=gather_pad,
        )
        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,
        attention_mode: str = "xformers",
        use_rope: bool = False,
        rope_scaling: Optional[Dict] = None,
        compress_kv_factor: Optional[Tuple] = None,
        block_idx: Optional[int] = None,
    ):
        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"

        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)
        else:
            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)

        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 if only_cross_attention else None,
            upcast_attention=upcast_attention,
            attention_mode=attention_mode,
            use_rope=use_rope,
            rope_scaling=rope_scaling,
            compress_kv_factor=compress_kv_factor,
        )

        # 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.
            self.norm2 = (
                AdaLayerNorm(dim, num_embeds_ada_norm)
                if self.use_ada_layer_norm
                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
            )
            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,
                attention_mode=attention_mode,  # only xformers support attention_mask
                use_rope=False,  # do not position in cross attention
                compress_kv_factor=None,
            )  # is self-attn if encoder_hidden_states is none
        else:
            self.norm2 = None
            self.attn2 = None

        # 3. Feed-forward
        if not self.use_ada_layer_norm_single:
            self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)

        self.ff = FeedForward(
            dim,
            dropout=dropout,
            activation_fn=activation_fn,
            final_dropout=final_dropout,
        )

        # 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

        # pab
        self.cross_last = None
        self.cross_count = 0
        self.spatial_last = None
        self.spatial_count = 0
        self.block_idx = block_idx
        self.spatila_mlp_count = 0

    def set_cross_last(self, last_out: torch.Tensor):
        self.cross_last = last_out

    def set_spatial_last(self, last_out: torch.Tensor):
        self.spatial_last = last_out

    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,
        position_q: Optional[torch.LongTensor] = None,
        position_k: Optional[torch.LongTensor] = None,
        hw: Tuple[int, int] = None,
        org_timestep: Optional[torch.LongTensor] = None,
        all_timesteps=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]

        # 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)

        broadcast_spatial, self.spatial_count = if_broadcast_spatial(
            int(org_timestep[0]), self.spatial_count, self.block_idx
        )
        if broadcast_spatial:
            attn_output = self.spatial_last
            assert 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)
        else:
            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_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)

            attn_output = self.attn1(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
                attention_mask=attention_mask,
                position_q=position_q,
                position_k=position_k,
                last_shape=hw,
                **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

            if enable_pab():
                self.set_spatial_last(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:
            broadcast_cross, self.cross_count = if_broadcast_cross(int(org_timestep[0]), self.cross_count)
            if broadcast_cross:
                hidden_states = hidden_states + self.cross_last
            else:
                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
                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,
                    position_q=None,  # cross attn do not need relative position
                    position_k=None,
                    last_shape=None,
                    **cross_attention_kwargs,
                )
                hidden_states = attn_output + hidden_states

                if enable_pab():
                    self.set_cross_last(attn_output)

        if enable_pab():
            broadcast_mlp, self.spatila_mlp_count, broadcast_next, broadcast_range = if_broadcast_mlp(
                int(org_timestep[0]),
                self.spatila_mlp_count,
                self.block_idx,
                all_timesteps.tolist(),
                is_temporal=False,
            )

        if enable_pab() and broadcast_mlp:
            ff_output = get_mlp_output(
                broadcast_range,
                timestep=int(org_timestep[0]),
                block_idx=self.block_idx,
                is_temporal=False,
            )
        else:
            # 4. Feed-forward
            if 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

            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

            if enable_pab() and broadcast_next:
                save_mlp_output(
                    timestep=int(org_timestep[0]),
                    block_idx=self.block_idx,
                    ff_output=ff_output,
                    is_temporal=False,
                )

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

        return hidden_states


class AdaLayerNormSingle(nn.Module):
    r"""
    Norm layer adaptive layer norm single (adaLN-single).

    As proposed in PixArt-Alpha (see: https://arxiv.org/abs/2310.00426; Section 2.3).

    Parameters:
        embedding_dim (`int`): The size of each embedding vector.
        use_additional_conditions (`bool`): To use additional conditions for normalization or not.
    """

    def __init__(self, embedding_dim: int, use_additional_conditions: bool = False):
        super().__init__()

        self.emb = CombinedTimestepSizeEmbeddings(
            embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions
        )

        self.silu = nn.SiLU()
        self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)

    def forward(
        self,
        timestep: torch.Tensor,
        added_cond_kwargs: Dict[str, torch.Tensor] = None,
        batch_size: int = None,
        hidden_dtype: Optional[torch.dtype] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        # No modulation happening here.
        embedded_timestep = self.emb(
            timestep, batch_size=batch_size, hidden_dtype=hidden_dtype, resolution=None, aspect_ratio=None
        )
        return self.linear(self.silu(embedded_timestep)), embedded_timestep


@dataclass
class Transformer3DModelOutput(BaseOutput):
    """
    The output of [`Transformer2DModel`].

    Args:
        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
            distributions for the unnoised latent pixels.
    """

    sample: torch.FloatTensor


class LatteT2V(ModelMixin, ConfigMixin):
    _supports_gradient_checkpointing = True

    """
    A 2D Transformer model for image-like data.

    Parameters:
        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
        in_channels (`int`, *optional*):
            The number of channels in the input and output (specify if the input is **continuous**).
        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
            This is fixed during training since it is used to learn a number of position embeddings.
        num_vector_embeds (`int`, *optional*):
            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
            Includes the class for the masked latent pixel.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
        num_embeds_ada_norm ( `int`, *optional*):
            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
            added to the hidden states.

            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
        attention_bias (`bool`, *optional*):
            Configure if the `TransformerBlocks` attention should contain a bias parameter.
    """

    @register_to_config
    def __init__(
        self,
        num_attention_heads: int = 16,
        patch_size_t: int = 1,
        attention_head_dim: int = 88,
        in_channels: Optional[int] = None,
        out_channels: Optional[int] = None,
        num_layers: int = 1,
        dropout: float = 0.0,
        norm_num_groups: int = 32,
        cross_attention_dim: Optional[int] = None,
        attention_bias: bool = False,
        sample_size: Optional[int] = None,
        num_vector_embeds: Optional[int] = None,
        patch_size: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        use_linear_projection: bool = False,
        only_cross_attention: bool = False,
        double_self_attention: bool = False,
        upcast_attention: bool = False,
        norm_type: str = "layer_norm",
        norm_elementwise_affine: bool = True,
        norm_eps: float = 1e-5,
        attention_type: str = "default",
        caption_channels: int = None,
        video_length: int = 16,
        attention_mode: str = "flash",
        use_rope: bool = False,
        model_max_length: int = 300,
        rope_scaling_type: str = "linear",
        compress_kv_factor: int = 1,
        interpolation_scale_1d: float = None,
    ):
        super().__init__()
        self.use_linear_projection = use_linear_projection
        self.num_attention_heads = num_attention_heads
        self.attention_head_dim = attention_head_dim
        inner_dim = num_attention_heads * attention_head_dim
        self.video_length = video_length
        self.use_rope = use_rope
        self.model_max_length = model_max_length
        self.compress_kv_factor = compress_kv_factor
        self.num_layers = num_layers
        self.config.hidden_size = model_max_length

        assert not (self.compress_kv_factor != 1 and use_rope), "Can not both enable compressing kv and using rope"

        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear

        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
        # Define whether input is continuous or discrete depending on configuration
        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
        self.is_input_vectorized = num_vector_embeds is not None
        # self.is_input_patches = in_channels is not None and patch_size is not None
        self.is_input_patches = True

        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
            deprecation_message = (
                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
            )
            deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
            norm_type = "ada_norm"

        # 2. Define input layers
        assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"

        self.height = sample_size[0]
        self.width = sample_size[1]

        self.patch_size = patch_size
        interpolation_scale_2d = self.config.sample_size[0] // 64  # => 64 (= 512 pixart) has interpolation scale 1
        interpolation_scale_2d = max(interpolation_scale_2d, 1)
        self.pos_embed = PatchEmbed(
            height=sample_size[0],
            width=sample_size[1],
            patch_size=patch_size,
            in_channels=in_channels,
            embed_dim=inner_dim,
            interpolation_scale=interpolation_scale_2d,
        )

        # define temporal positional embedding
        if interpolation_scale_1d is None:
            if self.config.video_length % 2 == 1:
                interpolation_scale_1d = (
                    self.config.video_length - 1
                ) // 16  # => 16 (= 16 Latte) has interpolation scale 1
            else:
                interpolation_scale_1d = self.config.video_length // 16  # => 16 (= 16 Latte) has interpolation scale 1
        # interpolation_scale_1d = self.config.video_length // 5  #
        interpolation_scale_1d = max(interpolation_scale_1d, 1)
        temp_pos_embed = get_1d_sincos_pos_embed(
            inner_dim, video_length, interpolation_scale=interpolation_scale_1d
        )  # 1152 hidden size
        self.register_buffer("temp_pos_embed", torch.from_numpy(temp_pos_embed).float().unsqueeze(0), persistent=False)

        rope_scaling = None
        if self.use_rope:
            self.position_getter_2d = PositionGetter2D()
            self.position_getter_1d = PositionGetter1D()
            rope_scaling = dict(
                type=rope_scaling_type, factor_2d=interpolation_scale_2d, factor_1d=interpolation_scale_1d
            )

        # 3. Define transformers blocks, spatial attention
        self.transformer_blocks = nn.ModuleList(
            [
                BasicTransformerBlock(
                    inner_dim,
                    num_attention_heads,
                    attention_head_dim,
                    dropout=dropout,
                    cross_attention_dim=cross_attention_dim,
                    activation_fn=activation_fn,
                    num_embeds_ada_norm=num_embeds_ada_norm,
                    attention_bias=attention_bias,
                    only_cross_attention=only_cross_attention,
                    double_self_attention=double_self_attention,
                    upcast_attention=upcast_attention,
                    norm_type=norm_type,
                    norm_elementwise_affine=norm_elementwise_affine,
                    norm_eps=norm_eps,
                    attention_type=attention_type,
                    attention_mode=attention_mode,
                    use_rope=use_rope,
                    rope_scaling=rope_scaling,
                    compress_kv_factor=(compress_kv_factor, compress_kv_factor)
                    if d >= num_layers // 2 and compress_kv_factor != 1
                    else None,  # follow pixart-sigma, apply in second-half layers
                    block_idx=d,
                )
                for d in range(num_layers)
            ]
        )

        # Define temporal transformers blocks
        self.temporal_transformer_blocks = nn.ModuleList(
            [
                BasicTransformerBlock_(  # one attention
                    inner_dim,
                    num_attention_heads,  # num_attention_heads
                    attention_head_dim,  # attention_head_dim 72
                    dropout=dropout,
                    cross_attention_dim=None,
                    activation_fn=activation_fn,
                    num_embeds_ada_norm=num_embeds_ada_norm,
                    attention_bias=attention_bias,
                    only_cross_attention=only_cross_attention,
                    double_self_attention=False,
                    upcast_attention=upcast_attention,
                    norm_type=norm_type,
                    norm_elementwise_affine=norm_elementwise_affine,
                    norm_eps=norm_eps,
                    attention_type=attention_type,
                    attention_mode=attention_mode,
                    use_rope=use_rope,
                    rope_scaling=rope_scaling,
                    compress_kv_factor=(compress_kv_factor,)
                    if d >= num_layers // 2 and compress_kv_factor != 1
                    else None,  # follow pixart-sigma, apply in second-half layers
                    block_idx=d,
                )
                for d in range(num_layers)
            ]
        )

        # 4. Define output layers
        self.out_channels = in_channels if out_channels is None else out_channels
        if self.is_input_continuous:
            # TODO: should use out_channels for continuous projections
            if use_linear_projection:
                self.proj_out = linear_cls(inner_dim, in_channels)
            else:
                self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
        elif self.is_input_vectorized:
            self.norm_out = nn.LayerNorm(inner_dim)
            self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
        elif self.is_input_patches and norm_type != "ada_norm_single":
            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
            self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
            self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
        elif self.is_input_patches and norm_type == "ada_norm_single":
            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
            self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
            self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)

        # 5. PixArt-Alpha blocks.
        self.adaln_single = None
        self.use_additional_conditions = False
        if norm_type == "ada_norm_single":
            # self.use_additional_conditions = self.config.sample_size[0] == 128  # False, 128 -> 1024
            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
            # additional conditions until we find better name
            self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)

        self.caption_projection = None
        if caption_channels is not None:
            self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)

        self.gradient_checkpointing = False

    def _set_gradient_checkpointing(self, module, value=False):
        self.gradient_checkpointing = value

    def make_position(self, b, t, use_image_num, h, w, device):
        pos_hw = self.position_getter_2d(b * (t + use_image_num), h, w, device)  # fake_b = b*(t+use_image_num)
        pos_t = self.position_getter_1d(b * h * w, t, device)  # fake_b = b*h*w
        return pos_hw, pos_t

    def make_attn_mask(self, attention_mask, frame, dtype):
        attention_mask = rearrange(attention_mask, "b t h w -> (b t) 1 (h w)")
        # assume that mask is expressed as:
        #   (1 = keep,      0 = discard)
        # convert mask into a bias that can be added to attention scores:
        #   (keep = +0,     discard = -10000.0)
        attention_mask = (1 - attention_mask.to(dtype)) * -10000.0
        attention_mask = attention_mask.to(self.dtype)
        return attention_mask

    def vae_to_diff_mask(self, attention_mask, use_image_num):
        dtype = attention_mask.dtype
        # b, t+use_image_num, h, w, assume t as channel
        # this version do not use 3d patch embedding
        attention_mask = F.max_pool2d(
            attention_mask, kernel_size=(self.patch_size, self.patch_size), stride=(self.patch_size, self.patch_size)
        )
        attention_mask = attention_mask.bool().to(dtype)
        return attention_mask

    def forward(
        self,
        hidden_states: torch.Tensor,
        timestep: Optional[torch.LongTensor] = None,
        all_timesteps=None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        added_cond_kwargs: Dict[str, torch.Tensor] = None,
        class_labels: Optional[torch.LongTensor] = None,
        cross_attention_kwargs: Dict[str, Any] = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        use_image_num: int = 0,
        enable_temporal_attentions: bool = True,
        return_dict: bool = True,
    ):
        """
        The [`Transformer2DModel`] forward method.

        Args:
            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, frame, channel, height, width)` if continuous):
                Input `hidden_states`.
            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
                self-attention.
            timestep ( `torch.LongTensor`, *optional*):
                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
                `AdaLayerZeroNorm`.
            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            attention_mask ( `torch.Tensor`, *optional*):
                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
                negative values to the attention scores corresponding to "discard" tokens.
            encoder_attention_mask ( `torch.Tensor`, *optional*):
                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:

                    * Mask `(batch, sequence_length)` True = keep, False = discard.
                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.

                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
                above. This bias will be added to the cross-attention scores.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
                tuple.

        Returns:
            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
            `tuple` where the first element is the sample tensor.
        """
        # 0. Split batch
        if get_cfg_parallel_size() > 1:
            (
                hidden_states,
                timestep,
                encoder_hidden_states,
                class_labels,
                attention_mask,
                encoder_attention_mask,
            ) = batch_func(
                partial(split_sequence, process_group=get_cfg_parallel_group(), dim=0),
                hidden_states,
                timestep,
                encoder_hidden_states,
                class_labels,
                attention_mask,
                encoder_attention_mask,
            )
        input_batch_size, c, frame, h, w = hidden_states.shape
        frame = frame - use_image_num  # 20-4=16
        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w").contiguous()
        org_timestep = timestep
        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
        # expects mask of shape:
        #   [batch, key_tokens]
        # adds singleton query_tokens dimension:
        #   [batch,                    1, key_tokens]
        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
        if attention_mask is None:
            attention_mask = torch.ones(
                (input_batch_size, frame + use_image_num, h, w), device=hidden_states.device, dtype=hidden_states.dtype
            )
        attention_mask = self.vae_to_diff_mask(attention_mask, use_image_num)
        dtype = attention_mask.dtype
        attention_mask_compress = F.max_pool2d(
            attention_mask.float(), kernel_size=self.compress_kv_factor, stride=self.compress_kv_factor
        )
        attention_mask_compress = attention_mask_compress.to(dtype)

        attention_mask = self.make_attn_mask(attention_mask, frame, hidden_states.dtype)
        attention_mask_compress = self.make_attn_mask(attention_mask_compress, frame, hidden_states.dtype)

        # 1 + 4, 1 -> video condition, 4 -> image condition
        # convert encoder_attention_mask to a bias the same way we do for attention_mask
        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:  # ndim == 2 means no image joint
            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
            encoder_attention_mask = repeat(encoder_attention_mask, "b 1 l -> (b f) 1 l", f=frame).contiguous()
            encoder_attention_mask = encoder_attention_mask.to(self.dtype)
        elif encoder_attention_mask is not None and encoder_attention_mask.ndim == 3:  # ndim == 3 means image joint
            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
            encoder_attention_mask_video = encoder_attention_mask[:, :1, ...]
            encoder_attention_mask_video = repeat(
                encoder_attention_mask_video, "b 1 l -> b (1 f) l", f=frame
            ).contiguous()
            encoder_attention_mask_image = encoder_attention_mask[:, 1:, ...]
            encoder_attention_mask = torch.cat([encoder_attention_mask_video, encoder_attention_mask_image], dim=1)
            encoder_attention_mask = rearrange(encoder_attention_mask, "b n l -> (b n) l").contiguous().unsqueeze(1)
            encoder_attention_mask = encoder_attention_mask.to(self.dtype)

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

        # 1. Input
        if self.is_input_patches:  # here
            height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
            hw = (height, width)
            num_patches = height * width

            hidden_states = self.pos_embed(hidden_states.to(self.dtype))  # alrady add positional embeddings

            if self.adaln_single is not None:
                if self.use_additional_conditions and added_cond_kwargs is None:
                    raise ValueError(
                        "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
                    )
                # batch_size = hidden_states.shape[0]
                batch_size = input_batch_size
                timestep, embedded_timestep = self.adaln_single(
                    timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
                )

        # 2. Blocks
        if self.caption_projection is not None:
            batch_size = hidden_states.shape[0]
            encoder_hidden_states = self.caption_projection(encoder_hidden_states.to(self.dtype))  # 3 120 1152

            if use_image_num != 0 and self.training:
                encoder_hidden_states_video = encoder_hidden_states[:, :1, ...]
                encoder_hidden_states_video = repeat(
                    encoder_hidden_states_video, "b 1 t d -> b (1 f) t d", f=frame
                ).contiguous()
                encoder_hidden_states_image = encoder_hidden_states[:, 1:, ...]
                encoder_hidden_states = torch.cat([encoder_hidden_states_video, encoder_hidden_states_image], dim=1)
                encoder_hidden_states_spatial = rearrange(encoder_hidden_states, "b f t d -> (b f) t d").contiguous()
            else:
                encoder_hidden_states_spatial = repeat(
                    encoder_hidden_states, "b 1 t d -> (b f) t d", f=frame
                ).contiguous()

        # prepare timesteps for spatial and temporal block
        timestep_spatial = repeat(timestep, "b d -> (b f) d", f=frame + use_image_num).contiguous()
        timestep_temp = repeat(timestep, "b d -> (b p) d", p=num_patches).contiguous()

        pos_hw, pos_t = None, None
        if self.use_rope:
            pos_hw, pos_t = self.make_position(
                input_batch_size, frame, use_image_num, height, width, hidden_states.device
            )

        if enable_sequence_parallel():
            set_temporal_pad(frame + use_image_num)
            set_spatial_pad(num_patches)
            hidden_states = self.split_from_second_dim(hidden_states, input_batch_size)
            encoder_hidden_states_spatial = self.split_from_second_dim(encoder_hidden_states_spatial, input_batch_size)
            timestep_spatial = self.split_from_second_dim(timestep_spatial, input_batch_size)
            attention_mask = self.split_from_second_dim(attention_mask, input_batch_size)
            attention_mask_compress = self.split_from_second_dim(attention_mask_compress, input_batch_size)
            temp_pos_embed = split_sequence(
                self.temp_pos_embed, get_sequence_parallel_group(), dim=1, grad_scale="down", pad=get_temporal_pad()
            )
        else:
            temp_pos_embed = self.temp_pos_embed

        for i, (spatial_block, temp_block) in enumerate(zip(self.transformer_blocks, self.temporal_transformer_blocks)):
            if self.training and self.gradient_checkpointing:
                hidden_states = torch.utils.checkpoint.checkpoint(
                    spatial_block,
                    hidden_states,
                    attention_mask_compress if i >= self.num_layers // 2 else attention_mask,
                    encoder_hidden_states_spatial,
                    encoder_attention_mask,
                    timestep_spatial,
                    cross_attention_kwargs,
                    class_labels,
                    pos_hw,
                    pos_hw,
                    hw,
                    use_reentrant=False,
                )

                if enable_temporal_attentions:
                    hidden_states = rearrange(hidden_states, "(b f) t d -> (b t) f d", b=input_batch_size).contiguous()

                    if use_image_num != 0:  # image-video joitn training
                        hidden_states_video = hidden_states[:, :frame, ...]
                        hidden_states_image = hidden_states[:, frame:, ...]

                        # if i == 0 and not self.use_rope:
                        if i == 0:
                            hidden_states_video = hidden_states_video + temp_pos_embed

                        hidden_states_video = torch.utils.checkpoint.checkpoint(
                            temp_block,
                            hidden_states_video,
                            None,  # attention_mask
                            None,  # encoder_hidden_states
                            None,  # encoder_attention_mask
                            timestep_temp,
                            cross_attention_kwargs,
                            class_labels,
                            pos_t,
                            pos_t,
                            (frame,),
                            use_reentrant=False,
                        )

                        hidden_states = torch.cat([hidden_states_video, hidden_states_image], dim=1)
                        hidden_states = rearrange(
                            hidden_states, "(b t) f d -> (b f) t d", b=input_batch_size
                        ).contiguous()

                    else:
                        # if i == 0 and not self.use_rope:
                        if i == 0:
                            hidden_states = hidden_states + temp_pos_embed

                        hidden_states = torch.utils.checkpoint.checkpoint(
                            temp_block,
                            hidden_states,
                            None,  # attention_mask
                            None,  # encoder_hidden_states
                            None,  # encoder_attention_mask
                            timestep_temp,
                            cross_attention_kwargs,
                            class_labels,
                            pos_t,
                            pos_t,
                            (frame,),
                            use_reentrant=False,
                        )

                        hidden_states = rearrange(
                            hidden_states, "(b t) f d -> (b f) t d", b=input_batch_size
                        ).contiguous()
            else:
                hidden_states = spatial_block(
                    hidden_states,
                    attention_mask_compress if i >= self.num_layers // 2 else attention_mask,
                    encoder_hidden_states_spatial,
                    encoder_attention_mask,
                    timestep_spatial,
                    cross_attention_kwargs,
                    class_labels,
                    pos_hw,
                    pos_hw,
                    hw,
                    org_timestep,
                    all_timesteps=all_timesteps,
                )

                if enable_temporal_attentions:
                    # b c f h w, f = 16 + 4
                    hidden_states = rearrange(hidden_states, "(b f) t d -> (b t) f d", b=input_batch_size).contiguous()

                    if use_image_num != 0 and self.training:
                        hidden_states_video = hidden_states[:, :frame, ...]
                        hidden_states_image = hidden_states[:, frame:, ...]

                        # if i == 0 and not self.use_rope:
                        #     hidden_states_video = hidden_states_video + temp_pos_embed

                        hidden_states_video = temp_block(
                            hidden_states_video,
                            None,  # attention_mask
                            None,  # encoder_hidden_states
                            None,  # encoder_attention_mask
                            timestep_temp,
                            cross_attention_kwargs,
                            class_labels,
                            pos_t,
                            pos_t,
                            (frame,),
                            org_timestep,
                        )

                        hidden_states = torch.cat([hidden_states_video, hidden_states_image], dim=1)
                        hidden_states = rearrange(
                            hidden_states, "(b t) f d -> (b f) t d", b=input_batch_size
                        ).contiguous()

                    else:
                        # if i == 0 and not self.use_rope:
                        if i == 0:
                            hidden_states = hidden_states + temp_pos_embed
                        hidden_states = temp_block(
                            hidden_states,
                            None,  # attention_mask
                            None,  # encoder_hidden_states
                            None,  # encoder_attention_mask
                            timestep_temp,
                            cross_attention_kwargs,
                            class_labels,
                            pos_t,
                            pos_t,
                            (frame,),
                            org_timestep,
                            all_timesteps=all_timesteps,
                        )

                        hidden_states = rearrange(
                            hidden_states, "(b t) f d -> (b f) t d", b=input_batch_size
                        ).contiguous()

        if enable_sequence_parallel():
            hidden_states = self.gather_from_second_dim(hidden_states, input_batch_size)

        if self.is_input_patches:
            if self.config.norm_type != "ada_norm_single":
                conditioning = self.transformer_blocks[0].norm1.emb(
                    timestep, class_labels, hidden_dtype=hidden_states.dtype
                )
                shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
                hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
                hidden_states = self.proj_out_2(hidden_states)
            elif self.config.norm_type == "ada_norm_single":
                embedded_timestep = repeat(embedded_timestep, "b d -> (b f) d", f=frame + use_image_num).contiguous()
                shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
                hidden_states = self.norm_out(hidden_states)
                # Modulation
                hidden_states = hidden_states * (1 + scale) + shift
                hidden_states = self.proj_out(hidden_states)

            # unpatchify
            if self.adaln_single is None:
                height = width = int(hidden_states.shape[1] ** 0.5)
            hidden_states = hidden_states.reshape(
                shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
            )
            hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
            output = hidden_states.reshape(
                shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
            )
            output = rearrange(output, "(b f) c h w -> b c f h w", b=input_batch_size).contiguous()

        # 3. Gather batch for data parallelism
        if get_cfg_parallel_size() > 1:
            output = gather_sequence(output, get_cfg_parallel_group(), dim=0)

        if not return_dict:
            return (output,)

        return Transformer3DModelOutput(sample=output)

    @classmethod
    def from_pretrained_2d(cls, pretrained_model_path, subfolder=None, **kwargs):
        if subfolder is not None:
            pretrained_model_path = os.path.join(pretrained_model_path, subfolder)

        config_file = os.path.join(pretrained_model_path, "config.json")
        if not os.path.isfile(config_file):
            raise RuntimeError(f"{config_file} does not exist")
        with open(config_file, "r") as f:
            config = json.load(f)

        model = cls.from_config(config, **kwargs)
        return model

    def split_from_second_dim(self, x, batch_size):
        x = x.view(batch_size, -1, *x.shape[1:])
        x = split_sequence(x, get_sequence_parallel_group(), dim=1, grad_scale="down", pad=get_temporal_pad())
        x = x.reshape(-1, *x.shape[2:])
        return x

    def gather_from_second_dim(self, x, batch_size):
        x = x.view(batch_size, -1, *x.shape[1:])
        x = gather_sequence(x, get_sequence_parallel_group(), dim=1, grad_scale="up", pad=get_temporal_pad())
        x = x.reshape(-1, *x.shape[2:])
        return x


# depth = num_layers * 2
def LatteT2V_XL_122(**kwargs):
    return LatteT2V(
        num_layers=28,
        attention_head_dim=72,
        num_attention_heads=16,
        patch_size_t=1,
        patch_size=2,
        norm_type="ada_norm_single",
        caption_channels=4096,
        cross_attention_dim=1152,
        **kwargs,
    )


def LatteT2V_D64_XL_122(**kwargs):
    return LatteT2V(
        num_layers=28,
        attention_head_dim=64,
        num_attention_heads=18,
        patch_size_t=1,
        patch_size=2,
        norm_type="ada_norm_single",
        caption_channels=4096,
        cross_attention_dim=1152,
        **kwargs,
    )


Latte_models = {
    "LatteT2V-XL/122": LatteT2V_XL_122,
    "LatteT2V-D64-XL/122": LatteT2V_D64_XL_122,
}