diffusers 0.23.0__py3-none-any.whl → 0.24.0__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

@@ -14,7 +14,12 @@
 
 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 = {}
@@ -23,6 +28,7 @@ if is_torch_available():
     _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
     _import_structure["autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
     _import_structure["autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
     _import_structure["autoencoder_tiny"] = ["AutoencoderTiny"]
     _import_structure["consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["controlnet"] = ["ControlNetModel"]
@@ -36,7 +42,9 @@ if is_torch_available():
     _import_structure["unet_2d"] = ["UNet2DModel"]
     _import_structure["unet_2d_condition"] = ["UNet2DConditionModel"]
     _import_structure["unet_3d_condition"] = ["UNet3DConditionModel"]
+    _import_structure["unet_kandi3"] = ["Kandinsky3UNet"]
     _import_structure["unet_motion_model"] = ["MotionAdapter", "UNetMotionModel"]
+    _import_structure["unet_spatio_temporal_condition"] = ["UNetSpatioTemporalConditionModel"]
     _import_structure["vq_model"] = ["VQModel"]
 
 if is_flax_available():
@@ -50,6 +58,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .adapter import MultiAdapter, T2IAdapter
         from .autoencoder_asym_kl import AsymmetricAutoencoderKL
         from .autoencoder_kl import AutoencoderKL
+        from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
         from .autoencoder_tiny import AutoencoderTiny
         from .consistency_decoder_vae import ConsistencyDecoderVAE
         from .controlnet import ControlNetModel
@@ -63,7 +72,9 @@ 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_kandi3 import Kandinsky3UNet
         from .unet_motion_model import MotionAdapter, UNetMotionModel
+        from .unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel
         from .vq_model import VQModel
 
     if is_flax_available():

diffusers/models/attention.py

@@ -25,6 +25,31 @@ from .lora import LoRACompatibleLinear
 from .normalization import AdaLayerNorm, AdaLayerNormZero
 
 
+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
 class GatedSelfAttentionDense(nn.Module):
     r"""
@@ -194,7 +219,12 @@ class BasicTransformerBlock(nn.Module):
         if not self.use_ada_layer_norm_single:
             self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
 
-        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,
+        )
 
         # 4. Fuser
         if attention_type == "gated" or attention_type == "gated-text-image":
@@ -208,7 +238,7 @@ 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
@@ -311,18 +341,8 @@ 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 +359,137 @@ 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, 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 FeedForward(nn.Module):
     r"""
     A feed-forward layer.

diffusers/models/attention_flax.py

@@ -110,7 +110,10 @@ def jax_memory_efficient_attention(
         )
 
     _, res = jax.lax.scan(
-        f=chunk_scanner, init=0, xs=None, length=math.ceil(num_q / query_chunk_size)  # start counter  # stop counter
+        f=chunk_scanner,
+        init=0,
+        xs=None,
+        length=math.ceil(num_q / query_chunk_size),  # start counter  # stop counter
     )
 
     return jnp.concatenate(res, axis=-3)  # fuse the chunked result back
@@ -138,6 +141,7 @@ class FlaxAttention(nn.Module):
             Parameters `dtype`
 
     """
+
     query_dim: int
     heads: int = 8
     dim_head: int = 64
@@ -262,6 +266,7 @@ class FlaxBasicTransformerBlock(nn.Module):
             Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
             enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
     """
+
     dim: int
     n_heads: int
     d_head: int
@@ -347,6 +352,7 @@ class FlaxTransformer2DModel(nn.Module):
             Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
             enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
     """
+
     in_channels: int
     n_heads: int
     d_head: int
@@ -442,6 +448,7 @@ class FlaxFeedForward(nn.Module):
         dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
             Parameters `dtype`
     """
+
     dim: int
     dropout: float = 0.0
     dtype: jnp.dtype = jnp.float32
@@ -471,6 +478,7 @@ class FlaxGEGLU(nn.Module):
         dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
             Parameters `dtype`
     """
+
     dim: int
     dropout: float = 0.0
     dtype: jnp.dtype = jnp.float32

diffusers/models/attention_processor.py

@@ -16,7 +16,7 @@ from typing import Callable, Optional, Union
 
 import torch
 import torch.nn.functional as F
-from torch import nn
+from torch import einsum, nn
 
 from ..utils import USE_PEFT_BACKEND, deprecate, logging
 from ..utils.import_utils import is_xformers_available
@@ -1975,6 +1975,288 @@ class LoRAAttnAddedKVProcessor(nn.Module):
         return attn.processor(attn, hidden_states, *args, **kwargs)
 
 
+class IPAdapterAttnProcessor(nn.Module):
+    r"""
+    Attention processor for IP-Adapater.
+
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        num_tokens (`int`, defaults to 4):
+            The context length of the image features.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, num_tokens=4, scale=1.0):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.num_tokens = num_tokens
+        self.scale = scale
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        scale=1.0,
+    ):
+        if scale != 1.0:
+            logger.warning("`scale` of IPAttnProcessor should be set with `set_ip_adapter_scale`.")
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        # split hidden states
+        end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+        encoder_hidden_states, ip_hidden_states = (
+            encoder_hidden_states[:, :end_pos, :],
+            encoder_hidden_states[:, end_pos:, :],
+        )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+
+        ip_key = attn.head_to_batch_dim(ip_key)
+        ip_value = attn.head_to_batch_dim(ip_value)
+
+        ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
+        ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
+        ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAdapterAttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        num_tokens (`int`, defaults to 4):
+            The context length of the image features.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, num_tokens=4, scale=1.0):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                f"{self.__class__.__name__} requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.num_tokens = num_tokens
+        self.scale = scale
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        scale=1.0,
+    ):
+        if scale != 1.0:
+            logger.warning("`scale` of IPAttnProcessor should be set by `set_ip_adapter_scale`.")
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        # split hidden states
+        end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+        encoder_hidden_states, ip_hidden_states = (
+            encoder_hidden_states[:, :end_pos, :],
+            encoder_hidden_states[:, end_pos:, :],
+        )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+
+        ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        ip_hidden_states = F.scaled_dot_product_attention(
+            query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+        )
+
+        ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        ip_hidden_states = ip_hidden_states.to(query.dtype)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+# TODO(Yiyi): This class should not exist, we can replace it with a normal attention processor I believe
+# this way torch.compile and co. will work as well
+class Kandi3AttnProcessor:
+    r"""
+    Default kandinsky3 proccesor for performing attention-related computations.
+    """
+
+    @staticmethod
+    def _reshape(hid_states, h):
+        b, n, f = hid_states.shape
+        d = f // h
+        return hid_states.unsqueeze(-1).reshape(b, n, h, d).permute(0, 2, 1, 3)
+
+    def __call__(
+        self,
+        attn,
+        x,
+        context,
+        context_mask=None,
+    ):
+        query = self._reshape(attn.to_q(x), h=attn.num_heads)
+        key = self._reshape(attn.to_k(context), h=attn.num_heads)
+        value = self._reshape(attn.to_v(context), h=attn.num_heads)
+
+        attention_matrix = einsum("b h i d, b h j d -> b h i j", query, key)
+
+        if context_mask is not None:
+            max_neg_value = -torch.finfo(attention_matrix.dtype).max
+            context_mask = context_mask.unsqueeze(1).unsqueeze(1)
+            attention_matrix = attention_matrix.masked_fill(~(context_mask != 0), max_neg_value)
+        attention_matrix = (attention_matrix * attn.scale).softmax(dim=-1)
+
+        out = einsum("b h i j, b h j d -> b h i d", attention_matrix, value)
+        out = out.permute(0, 2, 1, 3).reshape(out.shape[0], out.shape[2], -1)
+        out = attn.to_out[0](out)
+        return out
+
+
 LORA_ATTENTION_PROCESSORS = (
     LoRAAttnProcessor,
     LoRAAttnProcessor2_0,
@@ -1998,6 +2280,9 @@ CROSS_ATTENTION_PROCESSORS = (
     LoRAAttnProcessor,
     LoRAAttnProcessor2_0,
     LoRAXFormersAttnProcessor,
+    IPAdapterAttnProcessor,
+    IPAdapterAttnProcessor2_0,
+    Kandi3AttnProcessor,
 )
 
 AttentionProcessor = Union[