attention.py

# Copyright 2025 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.

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

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

from ..utils import deprecate, logging
from ..utils.import_utils import is_torch_npu_available, is_torch_xla_available, is_xformers_available
from ..utils.torch_utils import maybe_allow_in_graph
from .activations import GEGLU, GELU, ApproximateGELU, FP32SiLU, LinearActivation, SwiGLU
from .attention_processor import Attention, AttentionProcessor, JointAttnProcessor2_0
from .embeddings import SinusoidalPositionalEmbedding
from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm, SD35AdaLayerNormZeroX


if is_xformers_available():
    import xformers as xops
else:
    xops = None


logger = logging.get_logger(__name__)


class AttentionMixin:
    @property
    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

    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)

    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.
        """
        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.")

        for module in self.modules():
            if isinstance(module, AttentionModuleMixin):
                module.fuse_projections()

    def unfuse_qkv_projections(self):
        """Disables the fused QKV projection if enabled.

        > [!WARNING] > This API is 🧪 experimental.
        """
        for module in self.modules():
            if isinstance(module, AttentionModuleMixin):
                module.unfuse_projections()


class AttentionModuleMixin:
    _default_processor_cls = None
    _available_processors = []
    fused_projections = False

    def set_processor(self, processor: AttentionProcessor) -> None:
        """
        Set the attention processor to use.

        Args:
            processor (`AttnProcessor`):
                The attention processor to use.
        """
        # 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) -> "AttentionProcessor":
        """
        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

    def set_attention_backend(self, backend: str):
        from .attention_dispatch import AttentionBackendName

        available_backends = {x.value for x in AttentionBackendName.__members__.values()}
        if backend not in available_backends:
            raise ValueError(f"`{backend=}` must be one of the following: " + ", ".join(available_backends))

        backend = AttentionBackendName(backend.lower())
        self.processor._attention_backend = backend

    def set_use_npu_flash_attention(self, use_npu_flash_attention: bool) -> None:
        """
        Set whether to use NPU flash attention from `torch_npu` or not.

        Args:
            use_npu_flash_attention (`bool`): Whether to use NPU flash attention or not.
        """

        if use_npu_flash_attention:
            if not is_torch_npu_available():
                raise ImportError("torch_npu is not available")

        self.set_attention_backend("_native_npu")

    def set_use_xla_flash_attention(
        self,
        use_xla_flash_attention: bool,
        partition_spec: Optional[Tuple[Optional[str], ...]] = None,
        is_flux=False,
    ) -> None:
        """
        Set whether to use XLA flash attention from `torch_xla` or not.

        Args:
            use_xla_flash_attention (`bool`):
                Whether to use pallas flash attention kernel from `torch_xla` or not.
            partition_spec (`Tuple[]`, *optional*):
                Specify the partition specification if using SPMD. Otherwise None.
            is_flux (`bool`, *optional*, defaults to `False`):
                Whether the model is a Flux model.
        """
        if use_xla_flash_attention:
            if not is_torch_xla_available():
                raise ImportError("torch_xla is not available")

        self.set_attention_backend("_native_xla")

    def set_use_memory_efficient_attention_xformers(
        self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None
    ) -> None:
        """
        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`.
        """
        if use_memory_efficient_attention_xformers:
            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
                    if is_xformers_available():
                        dtype = None
                        if attention_op is not None:
                            op_fw, op_bw = attention_op
                            dtype, *_ = op_fw.SUPPORTED_DTYPES
                        q = torch.randn((1, 2, 40), device="cuda", dtype=dtype)
                        _ = xops.ops.memory_efficient_attention(q, q, q)
                except Exception as e:
                    raise e

                self.set_attention_backend("xformers")

    @torch.no_grad()
    def fuse_projections(self):
        """
        Fuse the query, key, and value projections into a single projection for efficiency.
        """
        # Skip if already fused
        if getattr(self, "fused_projections", False):
            return

        device = self.to_q.weight.data.device
        dtype = self.to_q.weight.data.dtype

        if hasattr(self, "is_cross_attention") and self.is_cross_attention:
            # Fuse cross-attention key-value projections
            concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data])
            in_features = concatenated_weights.shape[1]
            out_features = concatenated_weights.shape[0]

            self.to_kv = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
            self.to_kv.weight.copy_(concatenated_weights)
            if hasattr(self, "use_bias") and self.use_bias:
                concatenated_bias = torch.cat([self.to_k.bias.data, self.to_v.bias.data])
                self.to_kv.bias.copy_(concatenated_bias)
        else:
            # Fuse self-attention projections
            concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data])
            in_features = concatenated_weights.shape[1]
            out_features = concatenated_weights.shape[0]

            self.to_qkv = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
            self.to_qkv.weight.copy_(concatenated_weights)
            if hasattr(self, "use_bias") and self.use_bias:
                concatenated_bias = torch.cat([self.to_q.bias.data, self.to_k.bias.data, self.to_v.bias.data])
                self.to_qkv.bias.copy_(concatenated_bias)

        # Handle added projections for models like SD3, Flux, etc.
        if (
            getattr(self, "add_q_proj", None) is not None
            and getattr(self, "add_k_proj", None) is not None
            and getattr(self, "add_v_proj", None) is not None
        ):
            concatenated_weights = torch.cat(
                [self.add_q_proj.weight.data, self.add_k_proj.weight.data, self.add_v_proj.weight.data]
            )
            in_features = concatenated_weights.shape[1]
            out_features = concatenated_weights.shape[0]

            self.to_added_qkv = nn.Linear(
                in_features, out_features, bias=self.added_proj_bias, device=device, dtype=dtype
            )
            self.to_added_qkv.weight.copy_(concatenated_weights)
            if self.added_proj_bias:
                concatenated_bias = torch.cat(
                    [self.add_q_proj.bias.data, self.add_k_proj.bias.data, self.add_v_proj.bias.data]
                )
                self.to_added_qkv.bias.copy_(concatenated_bias)

        self.fused_projections = True

    @torch.no_grad()
    def unfuse_projections(self):
        """
        Unfuse the query, key, and value projections back to separate projections.
        """
        # Skip if not fused
        if not getattr(self, "fused_projections", False):
            return

        # Remove fused projection layers
        if hasattr(self, "to_qkv"):
            delattr(self, "to_qkv")

        if hasattr(self, "to_kv"):
            delattr(self, "to_kv")

        if hasattr(self, "to_added_qkv"):
            delattr(self, "to_added_qkv")

        self.fused_projections = False

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

        Args:
            slice_size (`int`):
                The slice size for attention computation.
        """
        if hasattr(self, "sliceable_head_dim") and 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}.")

        processor = None

        # Try to get a compatible processor for sliced attention
        if slice_size is not None:
            processor = self._get_compatible_processor("sliced")

        # If no processor was found or slice_size is None, use default processor
        if processor is None:
            processor = self.default_processor_cls()

        self.set_processor(processor)

    def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor:
        """
        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`.

        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:
        """
        Reshape the tensor for multi-head attention processing.

        Args:
            tensor (`torch.Tensor`): The tensor to reshape.
            out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor.

        Returns:
            `torch.Tensor`: The reshaped tensor.
        """
        head_size = self.heads
        if tensor.ndim == 3:
            batch_size, seq_len, dim = tensor.shape
            extra_dim = 1
        else:
            batch_size, extra_dim, seq_len, dim = tensor.shape
        tensor = tensor.reshape(batch_size, seq_len * extra_dim, 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 * extra_dim, dim // head_size)

        return tensor

    def get_attention_scores(
        self, query: torch.Tensor, key: torch.Tensor, attention_mask: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        """
        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.

        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:
        """
        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.
            batch_size (`int`): The batch size for repeating the attention mask.
            out_dim (`int`, *optional*, defaults to `3`): Output dimension.

        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:
        """
        Normalize the encoder hidden states.

        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 _chunked_feed_forward(ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int):
    # "feed_forward_chunk_size" can be used to save memory
    if hidden_states.shape[chunk_dim] % chunk_size != 0:
        raise ValueError(
            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
        )

    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
    ff_output = torch.cat(
        [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
        dim=chunk_dim,
    )
    return ff_output


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

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

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

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

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

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

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

        self.enabled = True

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

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

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

        return x


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

    Reference: https://huggingface.co/papers/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: int,
        num_attention_heads: int,
        attention_head_dim: int,
        context_pre_only: bool = False,
        qk_norm: Optional[str] = None,
        use_dual_attention: bool = False,
    ):
        super().__init__()

        self.use_dual_attention = use_dual_attention
        self.context_pre_only = context_pre_only
        context_norm_type = "ada_norm_continous" if context_pre_only else "ada_norm_zero"

        if use_dual_attention:
            self.norm1 = SD35AdaLayerNormZeroX(dim)
        else:
            self.norm1 = AdaLayerNormZero(dim)

        if context_norm_type == "ada_norm_continous":
            self.norm1_context = AdaLayerNormContinuous(
                dim, dim, elementwise_affine=False, eps=1e-6, bias=True, norm_type="layer_norm"
            )
        elif context_norm_type == "ada_norm_zero":
            self.norm1_context = AdaLayerNormZero(dim)
        else:
            raise ValueError(
                f"Unknown context_norm_type: {context_norm_type}, currently only support `ada_norm_continous`, `ada_norm_zero`"
            )

        if hasattr(F, "scaled_dot_product_attention"):
            processor = JointAttnProcessor2_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=context_pre_only,
            bias=True,
            processor=processor,
            qk_norm=qk_norm,
            eps=1e-6,
        )

        if use_dual_attention:
            self.attn2 = 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=qk_norm,
                eps=1e-6,
            )
        else:
            self.attn2 = None

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

        if not context_pre_only:
            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")
        else:
            self.norm2_context = None
            self.ff_context = None

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

    # Copied from diffusers.models.attention.BasicTransformerBlock.set_chunk_feed_forward
    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,
        encoder_hidden_states: torch.FloatTensor,
        temb: torch.FloatTensor,
        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        joint_attention_kwargs = joint_attention_kwargs or {}
        if self.use_dual_attention:
            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp, norm_hidden_states2, gate_msa2 = self.norm1(
                hidden_states, emb=temb
            )
        else:
            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)

        if self.context_pre_only:
            norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states, temb)
        else:
            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,
            **joint_attention_kwargs,
        )

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

        if self.use_dual_attention:
            attn_output2 = self.attn2(hidden_states=norm_hidden_states2, **joint_attention_kwargs)
            attn_output2 = gate_msa2.unsqueeze(1) * attn_output2
            hidden_states = hidden_states + attn_output2

        norm_hidden_states = self.norm2(hidden_states)
        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
        if self._chunk_size is not None:
            # "feed_forward_chunk_size" can be used to save memory
            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
        else:
            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`.
        if self.context_pre_only:
            encoder_hidden_states = None
        else:
            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]
            if self._chunk_size is not None:
                # "feed_forward_chunk_size" can be used to save memory
                context_ff_output = _chunked_feed_forward(
                    self.ff_context, norm_encoder_hidden_states, self._chunk_dim, self._chunk_size
                )
            else:
                context_ff_output = self.ff_context(norm_encoder_hidden_states)
            encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output

        return encoder_hidden_states, 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', 'ada_norm_continuous', 'layer_norm_i2vgen'
        norm_eps: float = 1e-5,
        final_dropout: bool = False,
        attention_type: str = "default",
        positional_embeddings: Optional[str] = None,
        num_positional_embeddings: Optional[int] = None,
        ada_norm_continous_conditioning_embedding_dim: Optional[int] = None,
        ada_norm_bias: Optional[int] = None,
        ff_inner_dim: Optional[int] = None,
        ff_bias: bool = True,
        attention_out_bias: bool = True,
    ):
        super().__init__()
        self.dim = dim
        self.num_attention_heads = num_attention_heads
        self.attention_head_dim = attention_head_dim
        self.dropout = dropout
        self.cross_attention_dim = cross_attention_dim
        self.activation_fn = activation_fn
        self.attention_bias = attention_bias
        self.double_self_attention = double_self_attention
        self.norm_elementwise_affine = norm_elementwise_affine
        self.positional_embeddings = positional_embeddings
        self.num_positional_embeddings = num_positional_embeddings
        self.only_cross_attention = only_cross_attention

        # We keep these boolean flags for backward-compatibility.
        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
        self.use_layer_norm = norm_type == "layer_norm"
        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"

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

        self.norm_type = norm_type
        self.num_embeds_ada_norm = num_embeds_ada_norm

        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 norm_type == "ada_norm":
            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
        elif norm_type == "ada_norm_zero":
            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
        elif norm_type == "ada_norm_continuous":
            self.norm1 = AdaLayerNormContinuous(
                dim,
                ada_norm_continous_conditioning_embedding_dim,
                norm_elementwise_affine,
                norm_eps,
                ada_norm_bias,
                "rms_norm",
            )
        else:
            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)

        self.attn1 = 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,
            out_bias=attention_out_bias,
        )

        # 2. Cross-Attn
        if cross_attention_dim is not None or double_self_attention:
            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
            # the second cross attention block.
            if norm_type == "ada_norm":
                self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
            elif norm_type == "ada_norm_continuous":
                self.norm2 = AdaLayerNormContinuous(
                    dim,
                    ada_norm_continous_conditioning_embedding_dim,
                    norm_elementwise_affine,
                    norm_eps,
                    ada_norm_bias,
                    "rms_norm",
                )
            else:
                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)

            self.attn2 = Attention(
                query_dim=dim,
                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                upcast_attention=upcast_attention,
                out_bias=attention_out_bias,
            )  # is self-attn if encoder_hidden_states is none
        else:
            if norm_type == "ada_norm_single":  # For Latte
                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
            else:
                self.norm2 = None
            self.attn2 = None

        # 3. Feed-forward
        if norm_type == "ada_norm_continuous":
            self.norm3 = AdaLayerNormContinuous(
                dim,
                ada_norm_continous_conditioning_embedding_dim,
                norm_elementwise_affine,
                norm_eps,
                ada_norm_bias,
                "layer_norm",
            )

        elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm"]:
            self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
        elif norm_type == "layer_norm_i2vgen":
            self.norm3 = None

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

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

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

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

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        timestep: Optional[torch.LongTensor] = None,
        cross_attention_kwargs: Dict[str, Any] = None,
        class_labels: Optional[torch.LongTensor] = None,
        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
    ) -> torch.Tensor:
        if cross_attention_kwargs is not None:
            if cross_attention_kwargs.get("scale", None) is not None:
                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")

        # Notice that normalization is always applied before the real computation in the following blocks.
        # 0. Self-Attention
        batch_size = hidden_states.shape[0]

        if self.norm_type == "ada_norm":
            norm_hidden_states = self.norm1(hidden_states, timestep)
        elif self.norm_type == "ada_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.norm_type in ["layer_norm", "layer_norm_i2vgen"]:
            norm_hidden_states = self.norm1(hidden_states)
        elif self.norm_type == "ada_norm_continuous":
            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
        elif self.norm_type == "ada_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
        else:
            raise ValueError("Incorrect norm used")

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

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

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

        if self.norm_type == "ada_norm_zero":
            attn_output = gate_msa.unsqueeze(1) * attn_output
        elif self.norm_type == "ada_norm_single":
            attn_output = gate_msa * attn_output

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

        # 1.2 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.norm_type == "ada_norm":
                norm_hidden_states = self.norm2(hidden_states, timestep)
            elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]:
                norm_hidden_states = self.norm2(hidden_states)
            elif self.norm_type == "ada_norm_single":
                # For PixArt norm2 isn't applied here:
                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
                norm_hidden_states = hidden_states
            elif self.norm_type == "ada_norm_continuous":
                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
            else:
                raise ValueError("Incorrect norm")

            if self.pos_embed is not None and self.norm_type != "ada_norm_single":
                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
        # i2vgen doesn't have this norm 🤷‍♂️
        if self.norm_type == "ada_norm_continuous":
            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
        elif not self.norm_type == "ada_norm_single":
            norm_hidden_states = self.norm3(hidden_states)

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

        if self.norm_type == "ada_norm_single":
            norm_hidden_states = self.norm2(hidden_states)
            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp

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

        if self.norm_type == "ada_norm_zero":
            ff_output = gate_mlp.unsqueeze(1) * ff_output
        elif self.norm_type == "ada_norm_single":
            ff_output = gate_mlp * ff_output

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

        return hidden_states


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

    Parameters:
        hidden_size (`int`):
            The dimensionality of the hidden layers in the model. This parameter determines the width of the model's
            hidden representations.
        intermediate_size (`int`): The intermediate dimension of the feedforward layer.
        multiple_of (`int`, *optional*): Value to ensure hidden dimension is a multiple
            of this value.
        ffn_dim_multiplier (float, *optional*): Custom multiplier for hidden
            dimension. Defaults to None.
    """

    def __init__(
        self,
        dim: int,
        inner_dim: int,
        multiple_of: Optional[int] = 256,
        ffn_dim_multiplier: Optional[float] = None,
    ):
        super().__init__()
        # custom hidden_size factor multiplier
        if ffn_dim_multiplier is not None:
            inner_dim = int(ffn_dim_multiplier * inner_dim)
        inner_dim = multiple_of * ((inner_dim + multiple_of - 1) // multiple_of)

        self.linear_1 = nn.Linear(
            dim,
            inner_dim,
            bias=False,
        )
        self.linear_2 = nn.Linear(
            inner_dim,
            dim,
            bias=False,
        )
        self.linear_3 = nn.Linear(
            dim,
            inner_dim,
            bias=False,
        )
        self.silu = FP32SiLU()

    def forward(self, x):
        return self.linear_2(self.silu(self.linear_1(x)) * self.linear_3(x))


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

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

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

        self.norm_in = nn.LayerNorm(dim)

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

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

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

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

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

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

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

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

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

        residual = hidden_states
        hidden_states = self.norm_in(hidden_states)

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

        if self.is_res:
            hidden_states = hidden_states + residual

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

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

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

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

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

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

        return hidden_states


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

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

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

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

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

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

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

        norm_hidden_states = self.norm1(hidden_states)

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

        hidden_states = attn_output + hidden_states

        norm_hidden_states = self.norm2(hidden_states)

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

        hidden_states = attn_output + hidden_states

        return hidden_states


@maybe_allow_in_graph
class FreeNoiseTransformerBlock(nn.Module):
    r"""
    A FreeNoise 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 (`int`, *optional*):
            The number of diffusion steps used during training. See `Transformer2DModel`.
        attention_bias (`bool`, defaults to `False`):
            Configure if the attentions should contain a bias parameter.
        only_cross_attention (`bool`, defaults to `False`):
            Whether to use only cross-attention layers. In this case two cross attention layers are used.
        double_self_attention (`bool`, defaults to `False`):
            Whether to use two self-attention layers. In this case no cross attention layers are used.
        upcast_attention (`bool`, defaults to `False`):
            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
        norm_elementwise_affine (`bool`, defaults to `True`):
            Whether to use learnable elementwise affine parameters for normalization.
        norm_type (`str`, defaults to `"layer_norm"`):
            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
        final_dropout (`bool` defaults to `False`):
            Whether to apply a final dropout after the last feed-forward layer.
        attention_type (`str`, defaults to `"default"`):
            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
        positional_embeddings (`str`, *optional*):
            The type of positional embeddings to apply to.
        num_positional_embeddings (`int`, *optional*, defaults to `None`):
            The maximum number of positional embeddings to apply.
        ff_inner_dim (`int`, *optional*):
            Hidden dimension of feed-forward MLP.
        ff_bias (`bool`, defaults to `True`):
            Whether or not to use bias in feed-forward MLP.
        attention_out_bias (`bool`, defaults to `True`):
            Whether or not to use bias in attention output project layer.
        context_length (`int`, defaults to `16`):
            The maximum number of frames that the FreeNoise block processes at once.
        context_stride (`int`, defaults to `4`):
            The number of frames to be skipped before starting to process a new batch of `context_length` frames.
        weighting_scheme (`str`, defaults to `"pyramid"`):
            The weighting scheme to use for weighting averaging of processed latent frames. As described in the
            Equation 9. of the [FreeNoise](https://huggingface.co/papers/2310.15169) paper, "pyramid" is the default
            setting used.
    """

    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        dropout: float = 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",
        norm_eps: float = 1e-5,
        final_dropout: bool = False,
        positional_embeddings: Optional[str] = None,
        num_positional_embeddings: Optional[int] = None,
        ff_inner_dim: Optional[int] = None,
        ff_bias: bool = True,
        attention_out_bias: bool = True,
        context_length: int = 16,
        context_stride: int = 4,
        weighting_scheme: str = "pyramid",
    ):
        super().__init__()
        self.dim = dim
        self.num_attention_heads = num_attention_heads
        self.attention_head_dim = attention_head_dim
        self.dropout = dropout
        self.cross_attention_dim = cross_attention_dim
        self.activation_fn = activation_fn
        self.attention_bias = attention_bias
        self.double_self_attention = double_self_attention
        self.norm_elementwise_affine = norm_elementwise_affine
        self.positional_embeddings = positional_embeddings
        self.num_positional_embeddings = num_positional_embeddings
        self.only_cross_attention = only_cross_attention

        self.set_free_noise_properties(context_length, context_stride, weighting_scheme)

        # We keep these boolean flags for backward-compatibility.
        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
        self.use_layer_norm = norm_type == "layer_norm"
        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"

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

        self.norm_type = norm_type
        self.num_embeds_ada_norm = num_embeds_ada_norm

        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
        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,
            out_bias=attention_out_bias,
        )

        # 2. Cross-Attn
        if cross_attention_dim is not None or double_self_attention:
            self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)

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

        # 3. Feed-forward
        self.ff = FeedForward(
            dim,
            dropout=dropout,
            activation_fn=activation_fn,
            final_dropout=final_dropout,
            inner_dim=ff_inner_dim,
            bias=ff_bias,
        )

        self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)

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

    def _get_frame_indices(self, num_frames: int) -> List[Tuple[int, int]]:
        frame_indices = []
        for i in range(0, num_frames - self.context_length + 1, self.context_stride):
            window_start = i
            window_end = min(num_frames, i + self.context_length)
            frame_indices.append((window_start, window_end))
        return frame_indices

    def _get_frame_weights(self, num_frames: int, weighting_scheme: str = "pyramid") -> List[float]:
        if weighting_scheme == "flat":
            weights = [1.0] * num_frames

        elif weighting_scheme == "pyramid":
            if num_frames % 2 == 0:
                # num_frames = 4 => [1, 2, 2, 1]
                mid = num_frames // 2
                weights = list(range(1, mid + 1))
                weights = weights + weights[::-1]
            else:
                # num_frames = 5 => [1, 2, 3, 2, 1]
                mid = (num_frames + 1) // 2
                weights = list(range(1, mid))
                weights = weights + [mid] + weights[::-1]

        elif weighting_scheme == "delayed_reverse_sawtooth":
            if num_frames % 2 == 0:
                # num_frames = 4 => [0.01, 2, 2, 1]
                mid = num_frames // 2
                weights = [0.01] * (mid - 1) + [mid]
                weights = weights + list(range(mid, 0, -1))
            else:
                # num_frames = 5 => [0.01, 0.01, 3, 2, 1]
                mid = (num_frames + 1) // 2
                weights = [0.01] * mid
                weights = weights + list(range(mid, 0, -1))
        else:
            raise ValueError(f"Unsupported value for weighting_scheme={weighting_scheme}")

        return weights

    def set_free_noise_properties(
        self, context_length: int, context_stride: int, weighting_scheme: str = "pyramid"
    ) -> None:
        self.context_length = context_length
        self.context_stride = context_stride
        self.weighting_scheme = weighting_scheme

    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0) -> None:
        # Sets chunk feed-forward
        self._chunk_size = chunk_size
        self._chunk_dim = dim

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        cross_attention_kwargs: Dict[str, Any] = None,
        *args,
        **kwargs,
    ) -> torch.Tensor:
        if cross_attention_kwargs is not None:
            if cross_attention_kwargs.get("scale", None) is not None:
                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")

        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}

        # hidden_states: [B x H x W, F, C]
        device = hidden_states.device
        dtype = hidden_states.dtype

        num_frames = hidden_states.size(1)
        frame_indices = self._get_frame_indices(num_frames)
        frame_weights = self._get_frame_weights(self.context_length, self.weighting_scheme)
        frame_weights = torch.tensor(frame_weights, device=device, dtype=dtype).unsqueeze(0).unsqueeze(-1)
        is_last_frame_batch_complete = frame_indices[-1][1] == num_frames

        # Handle out-of-bounds case if num_frames isn't perfectly divisible by context_length
        # For example, num_frames=25, context_length=16, context_stride=4, then we expect the ranges:
        #    [(0, 16), (4, 20), (8, 24), (10, 26)]
        if not is_last_frame_batch_complete:
            if num_frames < self.context_length:
                raise ValueError(f"Expected {num_frames=} to be greater or equal than {self.context_length=}")
            last_frame_batch_length = num_frames - frame_indices[-1][1]
            frame_indices.append((num_frames - self.context_length, num_frames))

        num_times_accumulated = torch.zeros((1, num_frames, 1), device=device)
        accumulated_values = torch.zeros_like(hidden_states)

        for i, (frame_start, frame_end) in enumerate(frame_indices):
            # The reason for slicing here is to ensure that if (frame_end - frame_start) is to handle
            # cases like frame_indices=[(0, 16), (16, 20)], if the user provided a video with 19 frames, or
            # essentially a non-multiple of `context_length`.
            weights = torch.ones_like(num_times_accumulated[:, frame_start:frame_end])
            weights *= frame_weights

            hidden_states_chunk = hidden_states[:, frame_start:frame_end]

            # Notice that normalization is always applied before the real computation in the following blocks.
            # 1. Self-Attention
            norm_hidden_states = self.norm1(hidden_states_chunk)

            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,
                **cross_attention_kwargs,
            )

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

            # 2. Cross-Attention
            if self.attn2 is not None:
                norm_hidden_states = self.norm2(hidden_states_chunk)

                if self.pos_embed is not None and self.norm_type != "ada_norm_single":
                    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_chunk = attn_output + hidden_states_chunk

            if i == len(frame_indices) - 1 and not is_last_frame_batch_complete:
                accumulated_values[:, -last_frame_batch_length:] += (
                    hidden_states_chunk[:, -last_frame_batch_length:] * weights[:, -last_frame_batch_length:]
                )
                num_times_accumulated[:, -last_frame_batch_length:] += weights[:, -last_frame_batch_length]
            else:
                accumulated_values[:, frame_start:frame_end] += hidden_states_chunk * weights
                num_times_accumulated[:, frame_start:frame_end] += weights

        # TODO(aryan): Maybe this could be done in a better way.
        #
        # Previously, this was:
        # hidden_states = torch.where(
        #    num_times_accumulated > 0, accumulated_values / num_times_accumulated, accumulated_values
        # )
        #
        # The reasoning for the change here is `torch.where` became a bottleneck at some point when golfing memory
        # spikes. It is particularly noticeable when the number of frames is high. My understanding is that this comes
        # from tensors being copied - which is why we resort to spliting and concatenating here. I've not particularly
        # looked into this deeply because other memory optimizations led to more pronounced reductions.
        hidden_states = torch.cat(
            [
                torch.where(num_times_split > 0, accumulated_split / num_times_split, accumulated_split)
                for accumulated_split, num_times_split in zip(
                    accumulated_values.split(self.context_length, dim=1),
                    num_times_accumulated.split(self.context_length, dim=1),
                )
            ],
            dim=1,
        ).to(dtype)

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

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

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

        return hidden_states


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

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

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

        if activation_fn == "gelu":
            act_fn = GELU(dim, inner_dim, bias=bias)
        if activation_fn == "gelu-approximate":
            act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
        elif activation_fn == "geglu":
            act_fn = GEGLU(dim, inner_dim, bias=bias)
        elif activation_fn == "geglu-approximate":
            act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
        elif activation_fn == "swiglu":
            act_fn = SwiGLU(dim, inner_dim, bias=bias)
        elif activation_fn == "linear-silu":
            act_fn = LinearActivation(dim, inner_dim, bias=bias, activation="silu")

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

    def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
        if len(args) > 0 or kwargs.get("scale", None) is not None:
            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
            deprecate("scale", "1.0.0", deprecation_message)
        for module in self.net:
            hidden_states = module(hidden_states)
        return hidden_states