# Copyright 2025 The SkyReels-V2 Team, The Wan Team and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import math
from typing import Any, Dict, Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F

from ...configuration_utils import ConfigMixin, register_to_config
from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
from ..attention import FeedForward
from ..attention_processor import Attention
from ..cache_utils import CacheMixin
from ..embeddings import (
    PixArtAlphaTextProjection,
    TimestepEmbedding,
    get_1d_rotary_pos_embed,
    get_1d_sincos_pos_embed_from_grid,
)
from ..modeling_outputs import Transformer2DModelOutput
from ..modeling_utils import ModelMixin, get_parameter_dtype
from ..normalization import FP32LayerNorm


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


class SkyReelsV2AttnProcessor2_0:
    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError(
                "SkyReelsV2AttnProcessor2_0 requires PyTorch 2.0. To use it, please upgrade PyTorch to 2.0."
            )

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        rotary_emb: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        encoder_hidden_states_img = None
        if attn.add_k_proj is not None:
            # 512 is the context length of the text encoder, hardcoded for now
            image_context_length = encoder_hidden_states.shape[1] - 512
            encoder_hidden_states_img = encoder_hidden_states[:, :image_context_length]
            encoder_hidden_states = encoder_hidden_states[:, image_context_length:]
        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states

        query = attn.to_q(hidden_states)
        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        if attn.norm_q is not None:
            query = attn.norm_q(query)
        if attn.norm_k is not None:
            key = attn.norm_k(key)

        query = query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
        key = key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
        value = value.unflatten(2, (attn.heads, -1)).transpose(1, 2)

        if rotary_emb is not None:

            def apply_rotary_emb(hidden_states: torch.Tensor, freqs: torch.Tensor):
                x_rotated = torch.view_as_complex(hidden_states.to(torch.float32).unflatten(3, (-1, 2)))
                x_out = torch.view_as_real(x_rotated * freqs).flatten(3, 4)
                return x_out.type_as(hidden_states)

            query = apply_rotary_emb(query, rotary_emb)
            key = apply_rotary_emb(key, rotary_emb)

        # I2V task
        hidden_states_img = None
        if encoder_hidden_states_img is not None:
            key_img = attn.add_k_proj(encoder_hidden_states_img)
            key_img = attn.norm_added_k(key_img)
            value_img = attn.add_v_proj(encoder_hidden_states_img)

            key_img = key_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
            value_img = value_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)

            hidden_states_img = F.scaled_dot_product_attention(
                query, key_img, value_img, attn_mask=None, dropout_p=0.0, is_causal=False
            )
            hidden_states_img = hidden_states_img.transpose(1, 2).flatten(2, 3)
            hidden_states_img = hidden_states_img.type_as(query)

        hidden_states = F.scaled_dot_product_attention(
            query,
            key,
            value,
            attn_mask=attention_mask,
            dropout_p=0.0,
            is_causal=False,
        )

        hidden_states = hidden_states.transpose(1, 2).flatten(2, 3)
        hidden_states = hidden_states.type_as(query)

        if hidden_states_img is not None:
            hidden_states = hidden_states + hidden_states_img

        hidden_states = attn.to_out[0](hidden_states)
        hidden_states = attn.to_out[1](hidden_states)
        return hidden_states


# Copied from diffusers.models.transformers.transformer_wan.WanImageEmbedding with WanImageEmbedding -> SkyReelsV2ImageEmbedding
class SkyReelsV2ImageEmbedding(torch.nn.Module):
    def __init__(self, in_features: int, out_features: int, pos_embed_seq_len=None):
        super().__init__()

        self.norm1 = FP32LayerNorm(in_features)
        self.ff = FeedForward(in_features, out_features, mult=1, activation_fn="gelu")
        self.norm2 = FP32LayerNorm(out_features)
        if pos_embed_seq_len is not None:
            self.pos_embed = nn.Parameter(torch.zeros(1, pos_embed_seq_len, in_features))
        else:
            self.pos_embed = None

    def forward(self, encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
        if self.pos_embed is not None:
            batch_size, seq_len, embed_dim = encoder_hidden_states_image.shape
            encoder_hidden_states_image = encoder_hidden_states_image.view(-1, 2 * seq_len, embed_dim)
            encoder_hidden_states_image = encoder_hidden_states_image + self.pos_embed

        hidden_states = self.norm1(encoder_hidden_states_image)
        hidden_states = self.ff(hidden_states)
        hidden_states = self.norm2(hidden_states)
        return hidden_states


class SkyReelsV2Timesteps(nn.Module):
    def __init__(self, num_channels: int, flip_sin_to_cos: bool, output_type: str = "pt"):
        super().__init__()
        self.num_channels = num_channels
        self.output_type = output_type
        self.flip_sin_to_cos = flip_sin_to_cos

    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
        original_shape = timesteps.shape
        t_emb = get_1d_sincos_pos_embed_from_grid(
            self.num_channels,
            timesteps,
            output_type=self.output_type,
            flip_sin_to_cos=self.flip_sin_to_cos,
        )
        # Reshape back to maintain batch structure
        if len(original_shape) > 1:
            t_emb = t_emb.reshape(*original_shape, self.num_channels)
        return t_emb


class SkyReelsV2TimeTextImageEmbedding(nn.Module):
    def __init__(
        self,
        dim: int,
        time_freq_dim: int,
        time_proj_dim: int,
        text_embed_dim: int,
        image_embed_dim: Optional[int] = None,
        pos_embed_seq_len: Optional[int] = None,
    ):
        super().__init__()

        self.timesteps_proj = SkyReelsV2Timesteps(num_channels=time_freq_dim, flip_sin_to_cos=True)
        self.time_embedder = TimestepEmbedding(in_channels=time_freq_dim, time_embed_dim=dim)
        self.act_fn = nn.SiLU()
        self.time_proj = nn.Linear(dim, time_proj_dim)
        self.text_embedder = PixArtAlphaTextProjection(text_embed_dim, dim, act_fn="gelu_tanh")

        self.image_embedder = None
        if image_embed_dim is not None:
            self.image_embedder = SkyReelsV2ImageEmbedding(image_embed_dim, dim, pos_embed_seq_len=pos_embed_seq_len)

    def forward(
        self,
        timestep: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        encoder_hidden_states_image: Optional[torch.Tensor] = None,
    ):
        timestep = self.timesteps_proj(timestep)

        time_embedder_dtype = get_parameter_dtype(self.time_embedder)
        if timestep.dtype != time_embedder_dtype and time_embedder_dtype != torch.int8:
            timestep = timestep.to(time_embedder_dtype)
        temb = self.time_embedder(timestep).type_as(encoder_hidden_states)
        timestep_proj = self.time_proj(self.act_fn(temb))

        encoder_hidden_states = self.text_embedder(encoder_hidden_states)
        if encoder_hidden_states_image is not None:
            encoder_hidden_states_image = self.image_embedder(encoder_hidden_states_image)

        return temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image


class SkyReelsV2RotaryPosEmbed(nn.Module):
    def __init__(
        self, attention_head_dim: int, patch_size: Tuple[int, int, int], max_seq_len: int, theta: float = 10000.0
    ):
        super().__init__()

        self.attention_head_dim = attention_head_dim
        self.patch_size = patch_size
        self.max_seq_len = max_seq_len

        h_dim = w_dim = 2 * (attention_head_dim // 6)
        t_dim = attention_head_dim - h_dim - w_dim

        freqs = []
        for dim in [t_dim, h_dim, w_dim]:
            freq = get_1d_rotary_pos_embed(
                dim, max_seq_len, theta, use_real=False, repeat_interleave_real=False, freqs_dtype=torch.float32
            )
            freqs.append(freq)
        self.freqs = torch.cat(freqs, dim=1)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, num_channels, num_frames, height, width = hidden_states.shape
        p_t, p_h, p_w = self.patch_size
        ppf, pph, ppw = num_frames // p_t, height // p_h, width // p_w

        freqs = self.freqs.to(hidden_states.device)
        freqs = freqs.split_with_sizes(
            [
                self.attention_head_dim // 2 - 2 * (self.attention_head_dim // 6),
                self.attention_head_dim // 6,
                self.attention_head_dim // 6,
            ],
            dim=1,
        )

        freqs_f = freqs[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
        freqs_h = freqs[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
        freqs_w = freqs[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
        freqs = torch.cat([freqs_f, freqs_h, freqs_w], dim=-1).reshape(1, 1, ppf * pph * ppw, -1)
        return freqs


class SkyReelsV2TransformerBlock(nn.Module):
    def __init__(
        self,
        dim: int,
        ffn_dim: int,
        num_heads: int,
        qk_norm: str = "rms_norm_across_heads",
        cross_attn_norm: bool = False,
        eps: float = 1e-6,
        added_kv_proj_dim: Optional[int] = None,
    ):
        super().__init__()

        # 1. Self-attention
        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
        self.attn1 = Attention(
            query_dim=dim,
            heads=num_heads,
            kv_heads=num_heads,
            dim_head=dim // num_heads,
            qk_norm=qk_norm,
            eps=eps,
            bias=True,
            cross_attention_dim=None,
            out_bias=True,
            processor=SkyReelsV2AttnProcessor2_0(),
        )

        # 2. Cross-attention
        self.attn2 = Attention(
            query_dim=dim,
            heads=num_heads,
            kv_heads=num_heads,
            dim_head=dim // num_heads,
            qk_norm=qk_norm,
            eps=eps,
            bias=True,
            cross_attention_dim=None,
            out_bias=True,
            added_kv_proj_dim=added_kv_proj_dim,
            added_proj_bias=True,
            processor=SkyReelsV2AttnProcessor2_0(),
        )
        self.norm2 = FP32LayerNorm(dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity()

        # 3. Feed-forward
        self.ffn = FeedForward(dim, inner_dim=ffn_dim, activation_fn="gelu-approximate")
        self.norm3 = FP32LayerNorm(dim, eps, elementwise_affine=False)

        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        temb: torch.Tensor,
        rotary_emb: torch.Tensor,
        attention_mask: torch.Tensor,
    ) -> torch.Tensor:
        if temb.dim() == 3:
            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
                self.scale_shift_table + temb.float()
            ).chunk(6, dim=1)
        elif temb.dim() == 4:
            # For 4D temb in Diffusion Forcing framework, we assume the shape is  (b, 6, f * pp_h * pp_w, inner_dim)
            e = (self.scale_shift_table.unsqueeze(2) + temb.float()).chunk(6, dim=1)
            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = [ei.squeeze(1) for ei in e]
        # 1. Self-attention
        norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
        attn_output = self.attn1(
            hidden_states=norm_hidden_states, rotary_emb=rotary_emb, attention_mask=attention_mask
        )
        hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
        # 2. Cross-attention
        norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
        attn_output = self.attn2(hidden_states=norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
        hidden_states = hidden_states + attn_output

        # 3. Feed-forward
        norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(
            hidden_states
        )
        ff_output = self.ffn(norm_hidden_states)
        hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
        return hidden_states


class SkyReelsV2Transformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
    r"""
    A Transformer model for video-like data used in the Wan-based SkyReels-V2 model.

    Args:
        patch_size (`Tuple[int]`, defaults to `(1, 2, 2)`):
            3D patch dimensions for video embedding (t_patch, h_patch, w_patch).
        num_attention_heads (`int`, defaults to `16`):
            Fixed length for text embeddings.
        attention_head_dim (`int`, defaults to `128`):
            The number of channels in each head.
        in_channels (`int`, defaults to `16`):
            The number of channels in the input.
        out_channels (`int`, defaults to `16`):
            The number of channels in the output.
        text_dim (`int`, defaults to `4096`):
            Input dimension for text embeddings.
        freq_dim (`int`, defaults to `256`):
            Dimension for sinusoidal time embeddings.
        ffn_dim (`int`, defaults to `8192`):
            Intermediate dimension in feed-forward network.
        num_layers (`int`, defaults to `32`):
            The number of layers of transformer blocks to use.
        window_size (`Tuple[int]`, defaults to `(-1, -1)`):
            Window size for local attention (-1 indicates global attention).
        cross_attn_norm (`bool`, defaults to `True`):
            Enable cross-attention normalization.
        qk_norm (`str`, *optional*, defaults to `"rms_norm_across_heads"`):
            Enable query/key normalization.
        eps (`float`, defaults to `1e-6`):
            Epsilon value for normalization layers.
        inject_sample_info (`bool`, defaults to `False`):
            Whether to inject sample information into the model.
        image_dim (`int`, *optional*):
            The dimension of the image embeddings.
        added_kv_proj_dim (`int`, *optional*):
            The dimension of the added key/value projection.
        rope_max_seq_len (`int`, defaults to `1024`):
            The maximum sequence length for the rotary embeddings.
        pos_embed_seq_len (`int`, *optional*):
            The sequence length for the positional embeddings.
    """

    _supports_gradient_checkpointing = True
    _skip_layerwise_casting_patterns = ["patch_embedding", "condition_embedder", "norm"]
    _no_split_modules = ["SkyReelsV2TransformerBlock"]
    _keep_in_fp32_modules = ["time_embedder", "scale_shift_table", "norm1", "norm2", "norm3"]
    _keys_to_ignore_on_load_unexpected = ["norm_added_q"]

    @register_to_config
    def __init__(
        self,
        patch_size: Tuple[int] = (1, 2, 2),
        num_attention_heads: int = 16,
        attention_head_dim: int = 128,
        in_channels: int = 16,
        out_channels: int = 16,
        text_dim: int = 4096,
        freq_dim: int = 256,
        ffn_dim: int = 8192,
        num_layers: int = 32,
        cross_attn_norm: bool = True,
        qk_norm: Optional[str] = "rms_norm_across_heads",
        eps: float = 1e-6,
        image_dim: Optional[int] = None,
        added_kv_proj_dim: Optional[int] = None,
        rope_max_seq_len: int = 1024,
        pos_embed_seq_len: Optional[int] = None,
        inject_sample_info: bool = False,
        num_frame_per_block: int = 1,
    ) -> None:
        super().__init__()

        inner_dim = num_attention_heads * attention_head_dim
        out_channels = out_channels or in_channels

        # 1. Patch & position embedding
        self.rope = SkyReelsV2RotaryPosEmbed(attention_head_dim, patch_size, rope_max_seq_len)
        self.patch_embedding = nn.Conv3d(in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)

        # 2. Condition embeddings
        # image_embedding_dim=1280 for I2V model
        self.condition_embedder = SkyReelsV2TimeTextImageEmbedding(
            dim=inner_dim,
            time_freq_dim=freq_dim,
            time_proj_dim=inner_dim * 6,
            text_embed_dim=text_dim,
            image_embed_dim=image_dim,
            pos_embed_seq_len=pos_embed_seq_len,
        )

        # 3. Transformer blocks
        self.blocks = nn.ModuleList(
            [
                SkyReelsV2TransformerBlock(
                    inner_dim, ffn_dim, num_attention_heads, qk_norm, cross_attn_norm, eps, added_kv_proj_dim
                )
                for _ in range(num_layers)
            ]
        )

        # 4. Output norm & projection
        self.norm_out = FP32LayerNorm(inner_dim, eps, elementwise_affine=False)
        self.proj_out = nn.Linear(inner_dim, out_channels * math.prod(patch_size))
        self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)

        if inject_sample_info:
            self.fps_embedding = nn.Embedding(2, inner_dim)
            self.fps_projection = FeedForward(inner_dim, inner_dim * 6, mult=1, activation_fn="linear-silu")

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        timestep: torch.LongTensor,
        encoder_hidden_states: torch.Tensor,
        encoder_hidden_states_image: Optional[torch.Tensor] = None,
        enable_diffusion_forcing: bool = False,
        fps: Optional[torch.Tensor] = None,
        return_dict: bool = True,
        attention_kwargs: Optional[Dict[str, Any]] = None,
    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
        if attention_kwargs is not None:
            attention_kwargs = attention_kwargs.copy()
            lora_scale = attention_kwargs.pop("scale", 1.0)
        else:
            lora_scale = 1.0

        if USE_PEFT_BACKEND:
            # weight the lora layers by setting `lora_scale` for each PEFT layer
            scale_lora_layers(self, lora_scale)
        else:
            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
                logger.warning(
                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
                )

        batch_size, num_channels, num_frames, height, width = hidden_states.shape
        p_t, p_h, p_w = self.config.patch_size
        post_patch_num_frames = num_frames // p_t
        post_patch_height = height // p_h
        post_patch_width = width // p_w

        rotary_emb = self.rope(hidden_states)

        hidden_states = self.patch_embedding(hidden_states)
        hidden_states = hidden_states.flatten(2).transpose(1, 2)

        causal_mask = None
        if self.config.num_frame_per_block > 1:
            block_num = post_patch_num_frames // self.config.num_frame_per_block
            range_tensor = torch.arange(block_num, device=hidden_states.device).repeat_interleave(
                self.config.num_frame_per_block
            )
            causal_mask = range_tensor.unsqueeze(0) <= range_tensor.unsqueeze(1)  # f, f
            causal_mask = causal_mask.view(post_patch_num_frames, 1, 1, post_patch_num_frames, 1, 1)
            causal_mask = causal_mask.repeat(
                1, post_patch_height, post_patch_width, 1, post_patch_height, post_patch_width
            )
            causal_mask = causal_mask.reshape(
                post_patch_num_frames * post_patch_height * post_patch_width,
                post_patch_num_frames * post_patch_height * post_patch_width,
            )
            causal_mask = causal_mask.unsqueeze(0).unsqueeze(0)

        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
            timestep, encoder_hidden_states, encoder_hidden_states_image
        )

        timestep_proj = timestep_proj.unflatten(-1, (6, -1))

        if encoder_hidden_states_image is not None:
            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)

        if self.config.inject_sample_info:
            fps = torch.tensor(fps, dtype=torch.long, device=hidden_states.device)

            fps_emb = self.fps_embedding(fps)
            if enable_diffusion_forcing:
                timestep_proj = timestep_proj + self.fps_projection(fps_emb).unflatten(1, (6, -1)).repeat(
                    timestep.shape[1], 1, 1
                )
            else:
                timestep_proj = timestep_proj + self.fps_projection(fps_emb).unflatten(1, (6, -1))

        if enable_diffusion_forcing:
            b, f = timestep.shape
            temb = temb.view(b, f, 1, 1, -1)
            timestep_proj = timestep_proj.view(b, f, 1, 1, 6, -1)  # (b, f, 1, 1, 6, inner_dim)
            temb = temb.repeat(1, 1, post_patch_height, post_patch_width, 1).flatten(1, 3)
            timestep_proj = timestep_proj.repeat(1, 1, post_patch_height, post_patch_width, 1, 1).flatten(
                1, 3
            )  # (b, f, pp_h, pp_w, 6, inner_dim) -> (b, f * pp_h * pp_w, 6, inner_dim)
            timestep_proj = timestep_proj.transpose(1, 2).contiguous()  # (b, 6, f * pp_h * pp_w, inner_dim)

        # 4. Transformer blocks
        if torch.is_grad_enabled() and self.gradient_checkpointing:
            for block in self.blocks:
                hidden_states = self._gradient_checkpointing_func(
                    block,
                    hidden_states,
                    encoder_hidden_states,
                    timestep_proj,
                    rotary_emb,
                    causal_mask,
                )
        else:
            for block in self.blocks:
                hidden_states = block(
                    hidden_states,
                    encoder_hidden_states,
                    timestep_proj,
                    rotary_emb,
                    causal_mask,
                )

        if temb.dim() == 2:
            # If temb is 2D, we assume it has time 1-D time embedding values for each batch.
            # For models:
            # - Skywork/SkyReels-V2-T2V-14B-540P-Diffusers
            # - Skywork/SkyReels-V2-T2V-14B-720P-Diffusers
            # - Skywork/SkyReels-V2-I2V-1.3B-540P-Diffusers
            # - Skywork/SkyReels-V2-I2V-14B-540P-Diffusers
            # - Skywork/SkyReels-V2-I2V-14B-720P-Diffusers
            shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
        elif temb.dim() == 3:
            # If temb is 3D, we assume it has 2-D time embedding values for each batch.
            # Each time embedding tensor includes values for each latent frame; thus Diffusion Forcing.
            # For models:
            # - Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers
            # - Skywork/SkyReels-V2-DF-14B-540P-Diffusers
            # - Skywork/SkyReels-V2-DF-14B-720P-Diffusers
            shift, scale = (self.scale_shift_table.unsqueeze(2) + temb.unsqueeze(1)).chunk(2, dim=1)
            shift, scale = shift.squeeze(1), scale.squeeze(1)

        # Move the shift and scale tensors to the same device as hidden_states.
        # When using multi-GPU inference via accelerate these will be on the
        # first device rather than the last device, which hidden_states ends up
        # on.
        shift = shift.to(hidden_states.device)
        scale = scale.to(hidden_states.device)

        hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)

        hidden_states = self.proj_out(hidden_states)

        hidden_states = hidden_states.reshape(
            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
        )
        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)

        if USE_PEFT_BACKEND:
            # remove `lora_scale` from each PEFT layer
            unscale_lora_layers(self, lora_scale)

        if not return_dict:
            return (output,)

        return Transformer2DModelOutput(sample=output)

    def _set_ar_attention(self, causal_block_size: int):
        self.register_to_config(num_frame_per_block=causal_block_size)