diffusers 0.23.1__py3-none-any.whl → 0.25.0__py3-none-any.whl

diffusers/loaders/utils.py

@@ -0,0 +1,59 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Dict
+
+import torch
+
+
+class AttnProcsLayers(torch.nn.Module):
+    def __init__(self, state_dict: Dict[str, torch.Tensor]):
+        super().__init__()
+        self.layers = torch.nn.ModuleList(state_dict.values())
+        self.mapping = dict(enumerate(state_dict.keys()))
+        self.rev_mapping = {v: k for k, v in enumerate(state_dict.keys())}
+
+        # .processor for unet, .self_attn for text encoder
+        self.split_keys = [".processor", ".self_attn"]
+
+        # we add a hook to state_dict() and load_state_dict() so that the
+        # naming fits with `unet.attn_processors`
+        def map_to(module, state_dict, *args, **kwargs):
+            new_state_dict = {}
+            for key, value in state_dict.items():
+                num = int(key.split(".")[1])  # 0 is always "layers"
+                new_key = key.replace(f"layers.{num}", module.mapping[num])
+                new_state_dict[new_key] = value
+
+            return new_state_dict
+
+        def remap_key(key, state_dict):
+            for k in self.split_keys:
+                if k in key:
+                    return key.split(k)[0] + k
+
+            raise ValueError(
+                f"There seems to be a problem with the state_dict: {set(state_dict.keys())}. {key} has to have one of {self.split_keys}."
+            )
+
+        def map_from(module, state_dict, *args, **kwargs):
+            all_keys = list(state_dict.keys())
+            for key in all_keys:
+                replace_key = remap_key(key, state_dict)
+                new_key = key.replace(replace_key, f"layers.{module.rev_mapping[replace_key]}")
+                state_dict[new_key] = state_dict[key]
+                del state_dict[key]
+
+        self._register_state_dict_hook(map_to)
+        self._register_load_state_dict_pre_hook(map_from, with_module=True)

diffusers/models/__init__.py

@@ -14,19 +14,26 @@
 
 from typing import TYPE_CHECKING
 
-from ..utils import DIFFUSERS_SLOW_IMPORT, _LazyModule, is_flax_available, is_torch_available
+from ..utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    _LazyModule,
+    is_flax_available,
+    is_torch_available,
+)
 
 
 _import_structure = {}
 
 if is_torch_available():
     _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
-    _import_structure["autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
-    _import_structure["autoencoder_kl"] = ["AutoencoderKL"]
-    _import_structure["autoencoder_tiny"] = ["AutoencoderTiny"]
-    _import_structure["consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
+    _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
+    _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
+    _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["controlnet"] = ["ControlNetModel"]
     _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
+    _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
     _import_structure["prior_transformer"] = ["PriorTransformer"]
     _import_structure["t5_film_transformer"] = ["T5FilmDecoder"]
@@ -36,7 +43,10 @@ if is_torch_available():
     _import_structure["unet_2d"] = ["UNet2DModel"]
     _import_structure["unet_2d_condition"] = ["UNet2DConditionModel"]
     _import_structure["unet_3d_condition"] = ["UNet3DConditionModel"]
+    _import_structure["unet_kandinsky3"] = ["Kandinsky3UNet"]
     _import_structure["unet_motion_model"] = ["MotionAdapter", "UNetMotionModel"]
+    _import_structure["unet_spatio_temporal_condition"] = ["UNetSpatioTemporalConditionModel"]
+    _import_structure["uvit_2d"] = ["UVit2DModel"]
     _import_structure["vq_model"] = ["VQModel"]
 
 if is_flax_available():
@@ -48,12 +58,16 @@ if is_flax_available():
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     if is_torch_available():
         from .adapter import MultiAdapter, T2IAdapter
-        from .autoencoder_asym_kl import AsymmetricAutoencoderKL
-        from .autoencoder_kl import AutoencoderKL
-        from .autoencoder_tiny import AutoencoderTiny
-        from .consistency_decoder_vae import ConsistencyDecoderVAE
+        from .autoencoders import (
+            AsymmetricAutoencoderKL,
+            AutoencoderKL,
+            AutoencoderKLTemporalDecoder,
+            AutoencoderTiny,
+            ConsistencyDecoderVAE,
+        )
         from .controlnet import ControlNetModel
         from .dual_transformer_2d import DualTransformer2DModel
+        from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin
         from .prior_transformer import PriorTransformer
         from .t5_film_transformer import T5FilmDecoder
@@ -63,7 +77,10 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .unet_2d import UNet2DModel
         from .unet_2d_condition import UNet2DConditionModel
         from .unet_3d_condition import UNet3DConditionModel
+        from .unet_kandinsky3 import Kandinsky3UNet
         from .unet_motion_model import MotionAdapter, UNetMotionModel
+        from .unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel
+        from .uvit_2d import UVit2DModel
         from .vq_model import VQModel
 
     if is_flax_available():

diffusers/models/activations.py

@@ -55,11 +55,12 @@ class GELU(nn.Module):
         dim_in (`int`): The number of channels in the input.
         dim_out (`int`): The number of channels in the output.
         approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
     """
 
-    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none"):
+    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True):
         super().__init__()
-        self.proj = nn.Linear(dim_in, dim_out)
+        self.proj = nn.Linear(dim_in, dim_out, bias=bias)
         self.approximate = approximate
 
     def gelu(self, gate: torch.Tensor) -> torch.Tensor:
@@ -81,13 +82,14 @@ class GEGLU(nn.Module):
     Parameters:
         dim_in (`int`): The number of channels in the input.
         dim_out (`int`): The number of channels in the output.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
     """
 
-    def __init__(self, dim_in: int, dim_out: int):
+    def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
         super().__init__()
         linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
 
-        self.proj = linear_cls(dim_in, dim_out * 2)
+        self.proj = linear_cls(dim_in, dim_out * 2, bias=bias)
 
     def gelu(self, gate: torch.Tensor) -> torch.Tensor:
         if gate.device.type != "mps":
@@ -109,11 +111,12 @@ class ApproximateGELU(nn.Module):
     Parameters:
         dim_in (`int`): The number of channels in the input.
         dim_out (`int`): The number of channels in the output.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
     """
 
-    def __init__(self, dim_in: int, dim_out: int):
+    def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
         super().__init__()
-        self.proj = nn.Linear(dim_in, dim_out)
+        self.proj = nn.Linear(dim_in, dim_out, bias=bias)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         x = self.proj(x)

diffusers/models/attention.py

@@ -14,6 +14,7 @@
 from typing import Any, Dict, Optional
 
 import torch
+import torch.nn.functional as F
 from torch import nn
 
 from ..utils import USE_PEFT_BACKEND
@@ -22,7 +23,32 @@ from .activations import GEGLU, GELU, ApproximateGELU
 from .attention_processor import Attention
 from .embeddings import SinusoidalPositionalEmbedding
 from .lora import LoRACompatibleLinear
-from .normalization import AdaLayerNorm, AdaLayerNormZero
+from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm
+
+
+def _chunked_feed_forward(
+    ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None
+):
+    # "feed_forward_chunk_size" can be used to save memory
+    if hidden_states.shape[chunk_dim] % chunk_size != 0:
+        raise ValueError(
+            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+        )
+
+    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
+    if lora_scale is None:
+        ff_output = torch.cat(
+            [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+            dim=chunk_dim,
+        )
+    else:
+        # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete
+        ff_output = torch.cat(
+            [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+            dim=chunk_dim,
+        )
+
+    return ff_output
 
 
 @maybe_allow_in_graph
@@ -123,6 +149,11 @@ class BasicTransformerBlock(nn.Module):
         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.only_cross_attention = only_cross_attention
@@ -131,6 +162,7 @@ class BasicTransformerBlock(nn.Module):
         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(
@@ -154,6 +186,15 @@ class BasicTransformerBlock(nn.Module):
             self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
         elif self.use_ada_layer_norm_zero:
             self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_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)
 
@@ -165,6 +206,7 @@ class BasicTransformerBlock(nn.Module):
             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
@@ -172,11 +214,20 @@ class BasicTransformerBlock(nn.Module):
             # 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 = (
-                AdaLayerNorm(dim, num_embeds_ada_norm)
-                if self.use_ada_layer_norm
-                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
-            )
+            if self.use_ada_layer_norm:
+                self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
+            elif self.use_ada_layer_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,
@@ -185,16 +236,33 @@ class BasicTransformerBlock(nn.Module):
                 dropout=dropout,
                 bias=attention_bias,
                 upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
             )  # is self-attn if encoder_hidden_states is none
         else:
             self.norm2 = None
             self.attn2 = None
 
         # 3. Feed-forward
-        if not self.use_ada_layer_norm_single:
-            self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+        if self.use_ada_layer_norm_continuous:
+            self.norm3 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "layer_norm",
+            )
+        elif not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
 
-        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
+        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":
@@ -208,7 +276,7 @@ class BasicTransformerBlock(nn.Module):
         self._chunk_size = None
         self._chunk_dim = 0
 
-    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
+    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
@@ -222,6 +290,7 @@ class BasicTransformerBlock(nn.Module):
         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.FloatTensor:
         # Notice that normalization is always applied before the real computation in the following blocks.
         # 0. Self-Attention
@@ -235,6 +304,8 @@ class BasicTransformerBlock(nn.Module):
             )
         elif self.use_layer_norm:
             norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_continuous:
+            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
         elif self.use_ada_layer_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)
@@ -284,6 +355,8 @@ class BasicTransformerBlock(nn.Module):
                 # 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.use_ada_layer_norm_continuous:
+                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
             else:
                 raise ValueError("Incorrect norm")
 
@@ -299,7 +372,9 @@ class BasicTransformerBlock(nn.Module):
             hidden_states = attn_output + hidden_states
 
         # 4. Feed-forward
-        if not self.use_ada_layer_norm_single:
+        if self.use_ada_layer_norm_continuous:
+            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif not self.use_ada_layer_norm_single:
             norm_hidden_states = self.norm3(hidden_states)
 
         if self.use_ada_layer_norm_zero:
@@ -311,18 +386,8 @@ class BasicTransformerBlock(nn.Module):
 
         if self._chunk_size is not None:
             # "feed_forward_chunk_size" can be used to save memory
-            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
-                raise ValueError(
-                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
-                )
-
-            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
-            ff_output = torch.cat(
-                [
-                    self.ff(hid_slice, scale=lora_scale)
-                    for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
-                ],
-                dim=self._chunk_dim,
+            ff_output = _chunked_feed_forward(
+                self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale
             )
         else:
             ff_output = self.ff(norm_hidden_states, scale=lora_scale)
@@ -339,6 +404,209 @@ class BasicTransformerBlock(nn.Module):
         return hidden_states
 
 
+@maybe_allow_in_graph
+class TemporalBasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block for video like data.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        time_mix_inner_dim (`int`): The number of channels for temporal attention.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        time_mix_inner_dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.is_res = dim == time_mix_inner_dim
+
+        self.norm_in = nn.LayerNorm(dim)
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.norm_in = nn.LayerNorm(dim)
+        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.FloatTensor,
+        num_frames: int,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        batch_frames, seq_length, channels = hidden_states.shape
+        batch_size = batch_frames // num_frames
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels)
+
+        residual = hidden_states
+        hidden_states = self.norm_in(hidden_states)
+
+        if self._chunk_size is not None:
+            hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            hidden_states = self.ff_in(hidden_states)
+
+        if self.is_res:
+            hidden_states = hidden_states + residual
+
+        norm_hidden_states = self.norm1(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+        hidden_states = attn_output + hidden_states
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = self.norm2(hidden_states)
+            attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.is_res:
+            hidden_states = ff_output + hidden_states
+        else:
+            hidden_states = ff_output
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels)
+
+        return hidden_states
+
+
+class SkipFFTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        kv_input_dim: int,
+        kv_input_dim_proj_use_bias: bool,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        if kv_input_dim != dim:
+            self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias)
+        else:
+            self.kv_mapper = None
+
+        self.norm1 = RMSNorm(dim, 1e-06)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim,
+            out_bias=attention_out_bias,
+        )
+
+        self.norm2 = RMSNorm(dim, 1e-06)
+
+        self.attn2 = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+        )
+
+    def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs):
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        if self.kv_mapper is not None:
+            encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states))
+
+        norm_hidden_states = self.norm1(hidden_states)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        norm_hidden_states = self.norm2(hidden_states)
+
+        attn_output = self.attn2(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        return hidden_states
+
+
 class FeedForward(nn.Module):
     r"""
     A feed-forward layer.
@@ -350,6 +618,7 @@ class FeedForward(nn.Module):
         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__(
@@ -360,20 +629,23 @@ class FeedForward(nn.Module):
         dropout: float = 0.0,
         activation_fn: str = "geglu",
         final_dropout: bool = False,
+        inner_dim=None,
+        bias: bool = True,
     ):
         super().__init__()
-        inner_dim = int(dim * mult)
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
         dim_out = dim_out if dim_out is not None else dim
         linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
 
         if activation_fn == "gelu":
-            act_fn = GELU(dim, inner_dim)
+            act_fn = GELU(dim, inner_dim, bias=bias)
         if activation_fn == "gelu-approximate":
-            act_fn = GELU(dim, inner_dim, approximate="tanh")
+            act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
         elif activation_fn == "geglu":
-            act_fn = GEGLU(dim, inner_dim)
+            act_fn = GEGLU(dim, inner_dim, bias=bias)
         elif activation_fn == "geglu-approximate":
-            act_fn = ApproximateGELU(dim, inner_dim)
+            act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
 
         self.net = nn.ModuleList([])
         # project in
@@ -381,7 +653,7 @@ class FeedForward(nn.Module):
         # project dropout
         self.net.append(nn.Dropout(dropout))
         # project out
-        self.net.append(linear_cls(inner_dim, dim_out))
+        self.net.append(linear_cls(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))