update

src/diffusers/models/attention.py

@@ -516,19 +516,46 @@ def _chunked_feed_forward(*args, **kwargs):
     return _actual_chunked_feed_forward(*args, **kwargs)
 
 
-class GatedSelfAttentionDense:
+class GatedSelfAttentionDense(nn.Module):
     r"""
-    Backward compatibility stub. Use transformers.modeling_common.GatedSelfAttentionDense instead.
+    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 __new__(cls, *args, **kwargs):
-        logger.warning(
-            "Importing `GatedSelfAttentionDense` from `diffusers.models.attention` is deprecated and will be removed in a future version. "
-            "Please use `from diffusers.models.transformers.modeling_common import GatedSelfAttentionDense` instead."
-        )
-        from .transformers.modeling_common import GatedSelfAttentionDense
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+        from .transformers.modeling_common import FeedForward
 
-        return GatedSelfAttentionDense(*args, **kwargs)
+        # we need a linear projection since we need cat visual feature and obj feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+        self.enabled = True
+
+    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
 
 
 class JointTransformerBlock:

src/diffusers/models/transformers/dit_transformer_2d.py

@@ -19,15 +19,218 @@ from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import logging
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..activations import GEGLU, GELU, ApproximateGELU
+from ..attention_processor import Attention
 from ..embeddings import PatchEmbed
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
-from .modeling_common import BasicTransformerBlock
+from ..normalization import AdaLayerNormZero
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
+class DiTFeedForward(nn.Module):
+    r"""
+    A feed-forward layer for DiT.
+
+    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)
+
+        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) -> torch.Tensor:
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class DiTTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block for DiT.
+
+    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.
+        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.
+        attention_bias (`bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        upcast_attention (`bool`, *optional*): Whether to upcast the attention computation to float32.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`): Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"ada_norm_zero"`): The normalization layer to use.
+        norm_eps (`float`, *optional*, defaults to 1e-5): A small constant added to the denominator in normalization layers.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "ada_norm_zero",
+        norm_eps: float = 1e-5,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.dropout = dropout
+        self.activation_fn = activation_fn
+        self.attention_bias = attention_bias
+        self.norm_elementwise_affine = norm_elementwise_affine
+
+        # Only support ada_norm_zero for DiT
+        if norm_type != "ada_norm_zero":
+            raise ValueError(f"DiTTransformerBlock only supports norm_type='ada_norm_zero', got {norm_type}")
+
+        if num_embeds_ada_norm is None:
+            raise ValueError("num_embeds_ada_norm must be provided for ada_norm_zero")
+
+        self.norm_type = norm_type
+        self.num_embeds_ada_norm = num_embeds_ada_norm
+
+        # 1. Self-Attention with AdaLayerNormZero
+        self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=None,
+            upcast_attention=upcast_attention,
+            out_bias=True,
+        )
+
+        # 2. Feed-forward
+        self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+        self.ff = DiTFeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=False,
+            inner_dim=None,
+            bias=True,
+        )
+
+        # 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,
+    ) -> torch.Tensor:
+        # 1. Self-Attention with adaptive norm
+        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+            hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+        )
+
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+        attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = attn_output + hidden_states
+
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2. Feed-forward
+        norm_hidden_states = self.norm3(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
+            if hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}."
+                )
+            num_chunks = hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [self.ff(hid_slice) for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+        hidden_states = ff_output + hidden_states
+
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
 class DiTTransformer2DModel(ModelMixin, ConfigMixin):
     r"""
     A 2D Transformer model as introduced in DiT (https://huggingface.co/papers/2212.09748).
@@ -121,7 +324,7 @@ class DiTTransformer2DModel(ModelMixin, ConfigMixin):
 
         self.transformer_blocks = nn.ModuleList(
             [
-                BasicTransformerBlock(
+                DiTTransformerBlock(
                     self.inner_dim,
                     self.config.num_attention_heads,
                     self.config.attention_head_dim,

src/diffusers/models/transformers/latte_transformer_3d.py

@@ -12,18 +12,362 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Optional
+from typing import Any, Dict, Optional
 
 import torch
 from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import logging
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import GatedSelfAttentionDense
+from ..attention_processor import Attention
 from ..cache_utils import CacheMixin
-from ..embeddings import PatchEmbed, PixArtAlphaTextProjection, get_1d_sincos_pos_embed_from_grid
+from ..embeddings import (
+    PatchEmbed,
+    PixArtAlphaTextProjection,
+    SinusoidalPositionalEmbedding,
+    get_1d_sincos_pos_embed_from_grid,
+)
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
-from ..normalization import AdaLayerNormSingle
-from .modeling_common import BasicTransformerBlock
+from ..normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormSingle, AdaLayerNormZero
+from .modeling_common import FeedForward, _chunked_feed_forward
+
+
+logger = logging.get_logger(__name__)
+
+
+@maybe_allow_in_graph
+class LatteTransformerBlock(nn.Module):
+    r"""
+    A Transformer block for Latte.
+
+    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",
+        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 LatteTransformer3DModel(ModelMixin, ConfigMixin, CacheMixin):
@@ -110,7 +454,7 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin, CacheMixin):
         # 2. Define spatial transformers blocks
         self.transformer_blocks = nn.ModuleList(
             [
-                BasicTransformerBlock(
+                LatteTransformerBlock(
                     inner_dim,
                     num_attention_heads,
                     attention_head_dim,
@@ -130,7 +474,7 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin, CacheMixin):
         # 3. Define temporal transformers blocks
         self.temporal_transformer_blocks = nn.ModuleList(
             [
-                BasicTransformerBlock(
+                LatteTransformerBlock(
                     inner_dim,
                     num_attention_heads,
                     attention_head_dim,

src/diffusers/models/transformers/modeling_common.py

@@ -21,6 +21,7 @@ import torch.nn.functional as F
 from ...utils import deprecate, logging
 from ...utils.torch_utils import maybe_allow_in_graph
 from ..activations import GEGLU, GELU, ApproximateGELU, FP32SiLU, LinearActivation, SwiGLU
+from ..attention import GatedSelfAttentionDense
 from ..attention_processor import Attention, JointAttnProcessor2_0
 from ..embeddings import SinusoidalPositionalEmbedding
 from ..normalization import (
@@ -50,48 +51,6 @@ def _chunked_feed_forward(ff: nn.Module, hidden_states: torch.Tensor, 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"""

src/diffusers/models/transformers/pixart_transformer_2d.py

@@ -18,12 +18,346 @@ from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import logging
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import GatedSelfAttentionDense
 from ..attention_processor import Attention, AttentionProcessor, AttnProcessor, FusedAttnProcessor2_0
-from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
+from ..embeddings import PatchEmbed, PixArtAlphaTextProjection, SinusoidalPositionalEmbedding
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
-from ..normalization import AdaLayerNormSingle
-from .modeling_common import BasicTransformerBlock
+from ..normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormSingle, AdaLayerNormZero
+from .modeling_common import FeedForward, _chunked_feed_forward
+
+
+@maybe_allow_in_graph
+class PixArtTransformerBlock(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
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -78,7 +412,7 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
     """
 
     _supports_gradient_checkpointing = True
-    _no_split_modules = ["BasicTransformerBlock", "PatchEmbed"]
+    _no_split_modules = ["PixArtTransformerBlock", "PatchEmbed"]
     _skip_layerwise_casting_patterns = ["pos_embed", "norm", "adaln_single"]
 
     @register_to_config
@@ -151,7 +485,7 @@ class PixArtTransformer2DModel(ModelMixin, ConfigMixin):
 
         self.transformer_blocks = nn.ModuleList(
             [
-                BasicTransformerBlock(
+                PixArtTransformerBlock(
                     self.inner_dim,
                     self.config.num_attention_heads,
                     self.config.attention_head_dim,

src/diffusers/models/transformers/prior_transformer.py

@@ -1,5 +1,5 @@
 from dataclasses import dataclass
-from typing import Dict, Optional, Union
+from typing import Any, Dict, Optional, Union
 
 import torch
 import torch.nn.functional as F
@@ -7,17 +7,356 @@ from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
-from ...utils import BaseOutput
+from ...utils import BaseOutput, logging
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import GatedSelfAttentionDense
 from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
+    Attention,
     AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
 )
-from ..embeddings import TimestepEmbedding, Timesteps
+from ..embeddings import SinusoidalPositionalEmbedding, TimestepEmbedding, Timesteps
 from ..modeling_utils import ModelMixin
-from .modeling_common import BasicTransformerBlock
+from ..normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero
+from .modeling_common import FeedForward, _chunked_feed_forward
+
+
+@maybe_allow_in_graph
+class PriorTransformerBlock(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
+
+
+logger = logging.get_logger(__name__)
 
 
 @dataclass
@@ -133,7 +472,7 @@ class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Pef
 
         self.transformer_blocks = nn.ModuleList(
             [
-                BasicTransformerBlock(
+                PriorTransformerBlock(
                     inner_dim,
                     num_attention_heads,
                     attention_head_dim,

src/diffusers/models/transformers/transformer_2d.py

@@ -19,16 +19,356 @@ from torch import nn
 
 from ...configuration_utils import LegacyConfigMixin, register_to_config
 from ...utils import deprecate, logging
-from ..embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import GatedSelfAttentionDense
+from ..attention_processor import Attention
+from ..embeddings import (
+    ImagePositionalEmbeddings,
+    PatchEmbed,
+    PixArtAlphaTextProjection,
+    SinusoidalPositionalEmbedding,
+)
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import LegacyModelMixin
-from ..normalization import AdaLayerNormSingle
-from .modeling_common import BasicTransformerBlock
+from ..normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormSingle, AdaLayerNormZero
+from .modeling_common import FeedForward, _chunked_feed_forward
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
+@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",
+        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 Transformer2DModelOutput(Transformer2DModelOutput):
     def __init__(self, *args, **kwargs):
         deprecation_message = "Importing `Transformer2DModelOutput` from `diffusers.models.transformer_2d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.modeling_outputs import Transformer2DModelOutput`, instead."

src/diffusers/models/transformers/transformer_sd3.py

@@ -15,6 +15,7 @@ from typing import Any, Dict, List, Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin, SD3Transformer2DLoadersMixin
@@ -29,13 +30,185 @@ from ..attention_processor import (
 from ..embeddings import CombinedTimestepTextProjEmbeddings, PatchEmbed
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
-from ..normalization import AdaLayerNormContinuous, AdaLayerNormZero
-from .modeling_common import FeedForward, JointTransformerBlock
+from ..normalization import AdaLayerNormContinuous, AdaLayerNormZero, SD35AdaLayerNormZeroX
+from .modeling_common import FeedForward, _chunked_feed_forward
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
+@maybe_allow_in_graph
+class SD3JointTransformerBlock(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,
+    ):
+        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 SD3SingleTransformerBlock(nn.Module):
     def __init__(
@@ -114,7 +287,7 @@ class SD3Transformer2DModel(
     """
 
     _supports_gradient_checkpointing = True
-    _no_split_modules = ["JointTransformerBlock"]
+    _no_split_modules = ["SD3JointTransformerBlock"]
     _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
 
     @register_to_config
@@ -155,7 +328,7 @@ class SD3Transformer2DModel(
 
         self.transformer_blocks = nn.ModuleList(
             [
-                JointTransformerBlock(
+                SD3JointTransformerBlock(
                     dim=self.inner_dim,
                     num_attention_heads=num_attention_heads,
                     attention_head_dim=attention_head_dim,

src/diffusers/models/transformers/transformer_temporal.py

@@ -18,11 +18,147 @@ import torch
 from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
-from ...utils import BaseOutput
+from ...utils import BaseOutput, logging
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention_processor import Attention
 from ..embeddings import TimestepEmbedding, Timesteps
 from ..modeling_utils import ModelMixin
 from ..resnet import AlphaBlender
-from .modeling_common import BasicTransformerBlock, TemporalBasicTransformerBlock
+from .modeling_common import FeedForward, _chunked_feed_forward
+from .transformer_2d import BasicTransformerBlock
+
+
+logger = logging.get_logger(__name__)
+
+
+@maybe_allow_in_graph
+class TemporalTransformerBlock(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
 
 
 @dataclass
@@ -101,7 +237,7 @@ class TransformerTemporalModel(ModelMixin, ConfigMixin):
         # 3. Define transformers blocks
         self.transformer_blocks = nn.ModuleList(
             [
-                BasicTransformerBlock(
+                TemporalTransformerBlock(
                     inner_dim,
                     num_attention_heads,
                     attention_head_dim,
@@ -254,7 +390,7 @@ class TransformerSpatioTemporalModel(nn.Module):
         time_mix_inner_dim = inner_dim
         self.temporal_transformer_blocks = nn.ModuleList(
             [
-                TemporalBasicTransformerBlock(
+                TemporalTransformerBlock(
                     inner_dim,
                     time_mix_inner_dim,
                     num_attention_heads,