transformer_flux.py

# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX 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.


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

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
from diffusers.models.attention import FeedForward
from diffusers.models.attention_processor import (
    Attention,
    AttentionProcessor,
    # FluxAttnProcessor2_0,
    FusedFluxAttnProcessor2_0,
)
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
from diffusers.utils.torch_utils import maybe_allow_in_graph
from diffusers.models.embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
from diffusers.models.modeling_outputs import Transformer2DModelOutput
import math


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

def gaussian_blur_2d(img, kernel_size, sigma):
    height = img.shape[-1]
    kernel_size = min(kernel_size, height - (height % 2 - 1))
    ksize_half = (kernel_size - 1) * 0.5

    x = torch.linspace(-ksize_half, ksize_half, steps=kernel_size)

    pdf = torch.exp(-0.5 * (x / sigma).pow(2))

    x_kernel = pdf / pdf.sum()
    x_kernel = x_kernel.to(device=img.device, dtype=img.dtype)

    kernel2d = torch.mm(x_kernel[:, None], x_kernel[None, :])
    kernel2d = kernel2d.expand(img.shape[-3], 1, kernel2d.shape[0], kernel2d.shape[1])

    padding = [kernel_size // 2, kernel_size // 2, kernel_size // 2, kernel_size // 2]

    img = F.pad(img, padding, mode="reflect")
    img = F.conv2d(img, kernel2d, groups=img.shape[-3])

    return img

class CustomFluxAttnProcessor2_0:
    """Attention processor used typically in processing the SD3-like self-attention projections."""

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: torch.FloatTensor = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        image_rotary_emb: Optional[torch.Tensor] = None,
        seg_blur_sigma=1.0,
        guidance_mode=None,
        seg_inf_blur_threshold=9999.0,
        apply_query_level_guidance=False, #whether to skip guidance for seg, pag.
        height = None,
        width=None,
        zip_weight=None,
    ) -> torch.FloatTensor:
        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape

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

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads
        
        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        ## TODO: chunk into 2 or 3 parts depending on just seg or cfg+seg
        if apply_query_level_guidance and guidance_mode == "seg":
            if encoder_hidden_states is None:
                query_enc_hid_states = query[:,:,:512,:]
                query = query[:,:,512:,:]
            query_org, query_ptb = query.chunk(2)
            query_ptb = query_ptb.permute(0, 1, 3, 2).view(batch_size//2, attn.heads * head_dim, height, width)

            if not (seg_blur_sigma > seg_inf_blur_threshold):
                kernel_size = math.ceil(6 * seg_blur_sigma) + 1 - math.ceil(6 * seg_blur_sigma) % 2
                query_ptb = gaussian_blur_2d(query_ptb, kernel_size, seg_blur_sigma)
            else:
                query_ptb[:] = query_ptb.mean(dim=(-2, -1), keepdim=True)

            query_ptb = query_ptb.view(batch_size//2, attn.heads, head_dim, height*width).permute(0,1,3,2)
            query = torch.cat((query_org, query_ptb), dim=0)
            if encoder_hidden_states is None:
                query = torch.cat([query_enc_hid_states, query], dim=2)


        if apply_query_level_guidance and guidance_mode == "cfg+seg":
            if encoder_hidden_states is None:
                query_enc_hid_states = query[:,:,:512,:]
                query = query[:,:,512:,:]
            query_org, query_ptb, query_uncond = query.chunk(3)
            query_ptb = query_ptb.permute(0, 1, 3, 2).view(batch_size//3, attn.heads * head_dim, height, width)
            
            if not (seg_blur_sigma > seg_inf_blur_threshold):
                kernel_size = math.ceil(6 * seg_blur_sigma) + 1 - math.ceil(6 * seg_blur_sigma) % 2
                query_ptb = gaussian_blur_2d(query_ptb, kernel_size, seg_blur_sigma)
            else:
                query_ptb[:] = query_ptb.mean(dim=(-2, -1), keepdim=True)

            query_ptb = query_ptb.view(batch_size//3, attn.heads, head_dim, height*width).permute(0,1,3,2)
            query = torch.cat((query_org, query_ptb, query_uncond), dim=0)
            if encoder_hidden_states is None:
                query = torch.cat([query_enc_hid_states, query], dim=2)

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

        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
        if encoder_hidden_states is not None:
            # `context` projections.
            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)

            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
                batch_size, -1, attn.heads, head_dim
            ).transpose(1, 2)
            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
                batch_size, -1, attn.heads, head_dim
            ).transpose(1, 2)
            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
                batch_size, -1, attn.heads, head_dim
            ).transpose(1, 2)

            if attn.norm_added_q is not None:
                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
            if attn.norm_added_k is not None:
                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)

            # attention
            query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)

        if image_rotary_emb is not None:
            from diffusers.models.embeddings import apply_rotary_emb

            query = apply_rotary_emb(query, image_rotary_emb)
            key = apply_rotary_emb(key, image_rotary_emb)

        if apply_query_level_guidance and guidance_mode == 'zip':
            attn_mask = torch.zeros(query.shape[0], query.shape[-2], key.shape[-2])
            #mask shape = bs * 2, l, s
            attn_mask = attn_mask + float('-inf')
            attn_mask = attn_mask.fill_diagonal_(0)
        else:
            attn_mask = None
            
        hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)

        if apply_query_level_guidance and guidance_mode == 'zip':
            hidden_states_ptb = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False, attn_mask=attn_mask)
            hidden_states = hidden_states_ptb + zip_weight * (hidden_states - hidden_states_ptb)
            print('query level attn')

        if apply_query_level_guidance and guidance_mode == 'pag':
            hidden_states_org, hidden_states_ptb = hidden_states.chunk(2)
            value_org, value_ptb = value.chunk(2)
            hidden_states = torch.cat([hidden_states_org, value_ptb])
        elif apply_query_level_guidance and guidance_mode == 'cfg+pag':
            hidden_states_org, hidden_states_ptb, hidden_states_uncond = hidden_states.chunk(3)
            value_org, value_ptb, value_uncond = value.chunk(3)
            hidden_states = torch.cat([hidden_states_org, value_ptb, hidden_states_uncond])
            
        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states = hidden_states.to(query.dtype)

        if encoder_hidden_states is not None:
            encoder_hidden_states, hidden_states = (
                hidden_states[:, : encoder_hidden_states.shape[1]],
                hidden_states[:, encoder_hidden_states.shape[1] :],
            )
            # linear proj
            hidden_states = attn.to_out[0](hidden_states)
            # dropout
            hidden_states = attn.to_out[1](hidden_states)
            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)

            return hidden_states, encoder_hidden_states
        else:
            return hidden_states
            
@maybe_allow_in_graph
class FluxSingleTransformerBlock(nn.Module):
    r"""
    A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.

    Reference: https://arxiv.org/abs/2403.03206

    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.
        context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
            processing of `context` conditions.
    """

    def __init__(self, dim, num_attention_heads, attention_head_dim, mlp_ratio=4.0):
        super().__init__()
        self.mlp_hidden_dim = int(dim * mlp_ratio)

        self.norm = AdaLayerNormZeroSingle(dim)
        self.proj_mlp = nn.Linear(dim, self.mlp_hidden_dim)
        self.act_mlp = nn.GELU(approximate="tanh")
        self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)

        processor = CustomFluxAttnProcessor2_0()
        self.attn = Attention(
            query_dim=dim,
            cross_attention_dim=None,
            dim_head=attention_head_dim,
            heads=num_attention_heads,
            out_dim=dim,
            bias=True,
            processor=processor,
            qk_norm="rms_norm",
            eps=1e-6,
            pre_only=True,
        )

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        temb: torch.FloatTensor,
        image_rotary_emb=None,
        guidance_kwargs=None
    ):
        residual = hidden_states
        norm_hidden_states, gate = self.norm(hidden_states, emb=temb)
        mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))

        attn_output = self.attn(
            hidden_states=norm_hidden_states,
            image_rotary_emb=image_rotary_emb,
            **guidance_kwargs
        )

        hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
        gate = gate.unsqueeze(1)
        hidden_states = gate * self.proj_out(hidden_states)
        hidden_states = residual + hidden_states
        if hidden_states.dtype == torch.float16:
            hidden_states = hidden_states.clip(-65504, 65504)

        return hidden_states


@maybe_allow_in_graph
class FluxTransformerBlock(nn.Module):
    r"""
    A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.

    Reference: https://arxiv.org/abs/2403.03206

    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.
        context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
            processing of `context` conditions.
    """

    def __init__(self, dim, num_attention_heads, attention_head_dim, qk_norm="rms_norm", eps=1e-6):
        super().__init__()

        self.norm1 = AdaLayerNormZero(dim)

        self.norm1_context = AdaLayerNormZero(dim)

        if hasattr(F, "scaled_dot_product_attention"):
            processor = CustomFluxAttnProcessor2_0()
        else:
            raise ValueError(
                "The current PyTorch version does not support the `scaled_dot_product_attention` function."
            )
        self.attn = Attention(
            query_dim=dim,
            cross_attention_dim=None,
            added_kv_proj_dim=dim,
            dim_head=attention_head_dim,
            heads=num_attention_heads,
            out_dim=dim,
            context_pre_only=False,
            bias=True,
            processor=processor,
            qk_norm=qk_norm,
            eps=eps,
        )

        self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
        self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")

        self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
        self.ff_context = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")

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

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: torch.FloatTensor,
        temb: torch.FloatTensor,
        image_rotary_emb=None,
        joint_attention_kwargs=None,
        guidance_kwargs=None
    ):
        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)

        norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
            encoder_hidden_states, emb=temb
        )

        # Attention.
        attn_output, context_attn_output = self.attn(
            hidden_states=norm_hidden_states,
            encoder_hidden_states=norm_encoder_hidden_states,
            image_rotary_emb=image_rotary_emb,
            **joint_attention_kwargs,
            **guidance_kwargs
        )

        # Process attention outputs for the `hidden_states`.
        attn_output = gate_msa.unsqueeze(1) * attn_output
        hidden_states = hidden_states + attn_output

        norm_hidden_states = self.norm2(hidden_states)
        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]

        ff_output = self.ff(norm_hidden_states)
        ff_output = gate_mlp.unsqueeze(1) * ff_output

        hidden_states = hidden_states + ff_output

        # Process attention outputs for the `encoder_hidden_states`.

        context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
        encoder_hidden_states = encoder_hidden_states + context_attn_output

        norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
        norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]

        context_ff_output = self.ff_context(norm_encoder_hidden_states)
        encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
        if encoder_hidden_states.dtype == torch.float16:
            encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)

        return encoder_hidden_states, hidden_states


class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
    """
    The Transformer model introduced in Flux.

    Reference: https://blackforestlabs.ai/announcing-black-forest-labs/

    Parameters:
        patch_size (`int`): Patch size to turn the input data into small patches.
        in_channels (`int`, *optional*, defaults to 16): The number of channels in the input.
        num_layers (`int`, *optional*, defaults to 18): The number of layers of MMDiT blocks to use.
        num_single_layers (`int`, *optional*, defaults to 18): The number of layers of single DiT blocks to use.
        attention_head_dim (`int`, *optional*, defaults to 64): The number of channels in each head.
        num_attention_heads (`int`, *optional*, defaults to 18): The number of heads to use for multi-head attention.
        joint_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
        pooled_projection_dim (`int`): Number of dimensions to use when projecting the `pooled_projections`.
        guidance_embeds (`bool`, defaults to False): Whether to use guidance embeddings.
    """

    _supports_gradient_checkpointing = True
    _no_split_modules = ["FluxTransformerBlock", "FluxSingleTransformerBlock"]

    @register_to_config
    def __init__(
        self,
        patch_size: int = 1,
        in_channels: int = 64,
        num_layers: int = 19,
        num_single_layers: int = 38,
        attention_head_dim: int = 128,
        num_attention_heads: int = 24,
        joint_attention_dim: int = 4096,
        pooled_projection_dim: int = 768,
        guidance_embeds: bool = False,
        axes_dims_rope: Tuple[int] = (16, 56, 56),
    ):
        super().__init__()
        self.out_channels = in_channels
        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim

        self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)

        text_time_guidance_cls = (
            CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
        )
        self.time_text_embed = text_time_guidance_cls(
            embedding_dim=self.inner_dim, pooled_projection_dim=self.config.pooled_projection_dim
        )

        self.context_embedder = nn.Linear(self.config.joint_attention_dim, self.inner_dim)
        self.x_embedder = torch.nn.Linear(self.config.in_channels, self.inner_dim)

        self.transformer_blocks = nn.ModuleList(
            [
                FluxTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=self.config.num_attention_heads,
                    attention_head_dim=self.config.attention_head_dim,
                )
                for i in range(self.config.num_layers)
            ]
        )

        self.single_transformer_blocks = nn.ModuleList(
            [
                FluxSingleTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=self.config.num_attention_heads,
                    attention_head_dim=self.config.attention_head_dim,
                )
                for i in range(self.config.num_single_layers)
            ]
        )

        self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
        self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)

        self.gradient_checkpointing = False

    @property
    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
    def attn_processors(self) -> Dict[str, AttentionProcessor]:
        r"""
        Returns:
            `dict` of attention processors: A dictionary containing all attention processors used in the model with
            indexed by its weight name.
        """
        # set recursively
        processors = {}

        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
            if hasattr(module, "get_processor"):
                processors[f"{name}.processor"] = module.get_processor()

            for sub_name, child in module.named_children():
                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

            return processors

        for name, module in self.named_children():
            fn_recursive_add_processors(name, module, processors)

        return processors

    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
        r"""
        Sets the attention processor to use to compute attention.

        Parameters:
            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
                The instantiated processor class or a dictionary of processor classes that will be set as the processor
                for **all** `Attention` layers.

                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
                processor. This is strongly recommended when setting trainable attention processors.

        """
        count = len(self.attn_processors.keys())

        if isinstance(processor, dict) and len(processor) != count:
            raise ValueError(
                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
            )

        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
            if hasattr(module, "set_processor"):
                if not isinstance(processor, dict):
                    module.set_processor(processor)
                else:
                    module.set_processor(processor.pop(f"{name}.processor"))

            for sub_name, child in module.named_children():
                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

        for name, module in self.named_children():
            fn_recursive_attn_processor(name, module, processor)

    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
    def fuse_qkv_projections(self):
        """
        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
        are fused. For cross-attention modules, key and value projection matrices are fused.

        <Tip warning={true}>

        This API is 🧪 experimental.

        </Tip>
        """
        self.original_attn_processors = None

        for _, attn_processor in self.attn_processors.items():
            if "Added" in str(attn_processor.__class__.__name__):
                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")

        self.original_attn_processors = self.attn_processors

        for module in self.modules():
            if isinstance(module, Attention):
                module.fuse_projections(fuse=True)

        self.set_attn_processor(FusedFluxAttnProcessor2_0())

    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
    def unfuse_qkv_projections(self):
        """Disables the fused QKV projection if enabled.

        <Tip warning={true}>

        This API is 🧪 experimental.

        </Tip>

        """
        if self.original_attn_processors is not None:
            self.set_attn_processor(self.original_attn_processors)

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor = None,
        pooled_projections: torch.Tensor = None,
        timestep: torch.LongTensor = None,
        img_ids: torch.Tensor = None,
        txt_ids: torch.Tensor = None,
        guidance: torch.Tensor = None,
        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
        controlnet_block_samples=None,
        controlnet_single_block_samples=None,
        return_dict: bool = True,
        guidance_kwargs=None
    ) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
        """
        The [`FluxTransformer2DModel`] forward method.

        Args:
            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
                Input `hidden_states`.
            encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
            pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
                from the embeddings of input conditions.
            timestep ( `torch.LongTensor`):
                Used to indicate denoising step.
            block_controlnet_hidden_states: (`list` of `torch.Tensor`):
                A list of tensors that if specified are added to the residuals of transformer blocks.
            joint_attention_kwargs (`dict`, *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).
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] 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.
        """
        if joint_attention_kwargs is not None:
            joint_attention_kwargs = joint_attention_kwargs.copy()
            lora_scale = joint_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 joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
                logger.warning(
                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
                )
        hidden_states = self.x_embedder(hidden_states)

        timestep = timestep.to(hidden_states.dtype) * 1000
        if guidance is not None:
            guidance = guidance.to(hidden_states.dtype) * 1000
        else:
            guidance = None
        temb = (
            self.time_text_embed(timestep, pooled_projections)
            if guidance is None
            else self.time_text_embed(timestep, guidance, pooled_projections)
        )
        encoder_hidden_states = self.context_embedder(encoder_hidden_states)

        if txt_ids.ndim == 3:
            logger.warning(
                "Passing `txt_ids` 3d torch.Tensor is deprecated."
                "Please remove the batch dimension and pass it as a 2d torch Tensor"
            )
            txt_ids = txt_ids[0]
        if img_ids.ndim == 3:
            logger.warning(
                "Passing `img_ids` 3d torch.Tensor is deprecated."
                "Please remove the batch dimension and pass it as a 2d torch Tensor"
            )
            img_ids = img_ids[0]

        ids = torch.cat((txt_ids, img_ids), dim=0)
        image_rotary_emb = self.pos_embed(ids)

        for index_block, block in enumerate(self.transformer_blocks):
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    hidden_states,
                    encoder_hidden_states,
                    temb,
                    image_rotary_emb,
                    **ckpt_kwargs,
                )

            else:
                guidance_kwargs['apply_query_level_guidance'] = index_block in guidance_kwargs.get('apply_transformer_blocks', [])
                encoder_hidden_states, hidden_states = block(
                    hidden_states=hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    temb=temb,
                    image_rotary_emb=image_rotary_emb,
                    joint_attention_kwargs=joint_attention_kwargs,
                    guidance_kwargs=guidance_kwargs
                )

            # controlnet residual
            if controlnet_block_samples is not None:
                interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
                interval_control = int(np.ceil(interval_control))
                hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]

        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)

        for index_block, block in enumerate(self.single_transformer_blocks):
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    hidden_states,
                    temb,
                    image_rotary_emb,
                    **ckpt_kwargs,
                )

            else:
                guidance_kwargs['apply_query_level_guidance'] = index_block in guidance_kwargs.get('apply_single_transformer_blocks', [])
                hidden_states = block(
                    hidden_states=hidden_states,
                    temb=temb,
                    image_rotary_emb=image_rotary_emb,
                    guidance_kwargs=guidance_kwargs
                )

            # controlnet residual
            if controlnet_single_block_samples is not None:
                interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples)
                interval_control = int(np.ceil(interval_control))
                hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
                    hidden_states[:, encoder_hidden_states.shape[1] :, ...]
                    + controlnet_single_block_samples[index_block // interval_control]
                )

        hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]

        hidden_states = self.norm_out(hidden_states, temb)
        output = self.proj_out(hidden_states)

        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)