diffusers 0.29.2__py3-none-any.whl → 0.30.0__py3-none-any.whl

diffusers/models/__init__.py

@@ -28,22 +28,32 @@ if is_torch_available():
     _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
     _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
     _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl_cogvideox"] = ["AutoencoderKLCogVideoX"]
     _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
+    _import_structure["autoencoders.autoencoder_oobleck"] = ["AutoencoderOobleck"]
     _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
     _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["autoencoders.vq_model"] = ["VQModel"]
     _import_structure["controlnet"] = ["ControlNetModel"]
+    _import_structure["controlnet_hunyuan"] = ["HunyuanDiT2DControlNetModel", "HunyuanDiT2DMultiControlNetModel"]
     _import_structure["controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
+    _import_structure["controlnet_sparsectrl"] = ["SparseControlNetModel"]
     _import_structure["controlnet_xs"] = ["ControlNetXSAdapter", "UNetControlNetXSModel"]
     _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
+    _import_structure["transformers.auraflow_transformer_2d"] = ["AuraFlowTransformer2DModel"]
+    _import_structure["transformers.cogvideox_transformer_3d"] = ["CogVideoXTransformer3DModel"]
     _import_structure["transformers.dit_transformer_2d"] = ["DiTTransformer2DModel"]
     _import_structure["transformers.dual_transformer_2d"] = ["DualTransformer2DModel"]
     _import_structure["transformers.hunyuan_transformer_2d"] = ["HunyuanDiT2DModel"]
+    _import_structure["transformers.latte_transformer_3d"] = ["LatteTransformer3DModel"]
+    _import_structure["transformers.lumina_nextdit2d"] = ["LuminaNextDiT2DModel"]
     _import_structure["transformers.pixart_transformer_2d"] = ["PixArtTransformer2DModel"]
     _import_structure["transformers.prior_transformer"] = ["PriorTransformer"]
+    _import_structure["transformers.stable_audio_transformer"] = ["StableAudioDiTModel"]
     _import_structure["transformers.t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
+    _import_structure["transformers.transformer_flux"] = ["FluxTransformer2DModel"]
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
     _import_structure["unets.unet_1d"] = ["UNet1DModel"]
@@ -69,23 +79,33 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .autoencoders import (
             AsymmetricAutoencoderKL,
             AutoencoderKL,
+            AutoencoderKLCogVideoX,
             AutoencoderKLTemporalDecoder,
+            AutoencoderOobleck,
             AutoencoderTiny,
             ConsistencyDecoderVAE,
             VQModel,
         )
         from .controlnet import ControlNetModel
+        from .controlnet_hunyuan import HunyuanDiT2DControlNetModel, HunyuanDiT2DMultiControlNetModel
         from .controlnet_sd3 import SD3ControlNetModel, SD3MultiControlNetModel
+        from .controlnet_sparsectrl import SparseControlNetModel
         from .controlnet_xs import ControlNetXSAdapter, UNetControlNetXSModel
         from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin
         from .transformers import (
+            AuraFlowTransformer2DModel,
+            CogVideoXTransformer3DModel,
             DiTTransformer2DModel,
             DualTransformer2DModel,
+            FluxTransformer2DModel,
             HunyuanDiT2DModel,
+            LatteTransformer3DModel,
+            LuminaNextDiT2DModel,
             PixArtTransformer2DModel,
             PriorTransformer,
             SD3Transformer2DModel,
+            StableAudioDiTModel,
             T5FilmDecoder,
             Transformer2DModel,
             TransformerTemporalModel,

diffusers/models/activations.py

@@ -123,6 +123,28 @@ class GEGLU(nn.Module):
             return hidden_states * self.gelu(gate)
 
 
+class SwiGLU(nn.Module):
+    r"""
+    A [variant](https://arxiv.org/abs/2002.05202) of the gated linear unit activation function. It's similar to `GEGLU`
+    but uses SiLU / Swish instead of GeLU.
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias)
+        self.activation = nn.SiLU()
+
+    def forward(self, hidden_states):
+        hidden_states = self.proj(hidden_states)
+        hidden_states, gate = hidden_states.chunk(2, dim=-1)
+        return hidden_states * self.activation(gate)
+
+
 class ApproximateGELU(nn.Module):
     r"""
     The approximate form of the Gaussian Error Linear Unit (GELU). For more details, see section 2 of this

diffusers/models/attention.py

@@ -11,7 +11,7 @@
 # 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
+from typing import Any, Dict, List, Optional, Tuple
 
 import torch
 import torch.nn.functional as F
@@ -19,7 +19,7 @@ from torch import nn
 
 from ..utils import deprecate, logging
 from ..utils.torch_utils import maybe_allow_in_graph
-from .activations import GEGLU, GELU, ApproximateGELU
+from .activations import GEGLU, GELU, ApproximateGELU, FP32SiLU, SwiGLU
 from .attention_processor import Attention, JointAttnProcessor2_0
 from .embeddings import SinusoidalPositionalEmbedding
 from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm
@@ -128,9 +128,9 @@ class JointTransformerBlock(nn.Module):
             query_dim=dim,
             cross_attention_dim=None,
             added_kv_proj_dim=dim,
-            dim_head=attention_head_dim // num_attention_heads,
+            dim_head=attention_head_dim,
             heads=num_attention_heads,
-            out_dim=attention_head_dim,
+            out_dim=dim,
             context_pre_only=context_pre_only,
             bias=True,
             processor=processor,
@@ -272,6 +272,17 @@ class BasicTransformerBlock(nn.Module):
         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.
@@ -359,7 +370,10 @@ class BasicTransformerBlock(nn.Module):
                 out_bias=attention_out_bias,
             )  # is self-attn if encoder_hidden_states is none
         else:
-            self.norm2 = None
+            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
@@ -373,7 +387,7 @@ class BasicTransformerBlock(nn.Module):
                 "layer_norm",
             )
 
-        elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm", "ada_norm_continuous"]:
+        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
@@ -439,7 +453,6 @@ class BasicTransformerBlock(nn.Module):
             ).chunk(6, dim=1)
             norm_hidden_states = self.norm1(hidden_states)
             norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
-            norm_hidden_states = norm_hidden_states.squeeze(1)
         else:
             raise ValueError("Incorrect norm used")
 
@@ -456,6 +469,7 @@ class BasicTransformerBlock(nn.Module):
             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":
@@ -527,6 +541,56 @@ class BasicTransformerBlock(nn.Module):
         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__()
+        inner_dim = int(2 * inner_dim / 3)
+        # 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"""
@@ -729,6 +793,319 @@ class SkipFFTransformerBlock(nn.Module):
         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://arxiv.org/abs/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 == "pyramid":
+            if num_frames % 2 == 0:
+                # num_frames = 4 => [1, 2, 2, 1]
+                weights = list(range(1, num_frames // 2 + 1))
+                weights = weights + weights[::-1]
+            else:
+                # num_frames = 5 => [1, 2, 3, 2, 1]
+                weights = list(range(1, num_frames // 2 + 1))
+                weights = weights + [num_frames // 2 + 1] + weights[::-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
+
+        hidden_states = torch.where(
+            num_times_accumulated > 0, accumulated_values / num_times_accumulated, accumulated_values
+        ).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.
@@ -767,6 +1144,8 @@ class FeedForward(nn.Module):
             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)
 
         self.net = nn.ModuleList([])
         # project in