unet_2d_condition.py

# Copyright 2024 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.
fromdataclassesimportdataclass
fromtypingimportAny, Dict, List, Optional, Tuple, Union

importtorch
importtorch.nnasnn
importtorch.utils.checkpoint

from ...configuration_utilsimportConfigMixin, register_to_config
from ...loadersimportPeftAdapterMixin, UNet2DConditionLoadersMixin
from ...loaders.single_file_modelimportFromOriginalModelMixin
from ...utilsimportUSE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
from ..activationsimportget_activation
from ..attention_processorimport (
ADDED_KV_ATTENTION_PROCESSORS,
CROSS_ATTENTION_PROCESSORS,
Attention,
AttentionProcessor,
AttnAddedKVProcessor,
AttnProcessor,
FusedAttnProcessor2_0,
)
from ..embeddingsimport (
GaussianFourierProjection,
GLIGENTextBoundingboxProjection,
ImageHintTimeEmbedding,
ImageProjection,
ImageTimeEmbedding,
TextImageProjection,
TextImageTimeEmbedding,
TextTimeEmbedding,
TimestepEmbedding,
Timesteps,
)
from ..modeling_utilsimportModelMixin
from .unet_2d_blocksimport (
get_down_block,
get_mid_block,
get_up_block,
)


logger=logging.get_logger(__name__) # pylint: disable=invalid-name


@dataclass
classUNet2DConditionOutput(BaseOutput):
"""
 The output of [`UNet2DConditionModel`].

 Args:
 sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)`):
 The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
 """

sample: torch.Tensor=None


classUNet2DConditionModel(
ModelMixin, ConfigMixin, FromOriginalModelMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin
):
r"""
 A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
 shaped output.

 This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
 for all models (such as downloading or saving).

 Parameters:
 sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
 Height and width of input/output sample.
 in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
 out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
 center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
 flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
 Whether to flip the sin to cos in the time embedding.
 freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
 down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
 The tuple of downsample blocks to use.
 mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
 Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
 `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
 up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
 The tuple of upsample blocks to use.
 only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
 Whether to include self-attention in the basic transformer blocks, see
 [`~models.attention.BasicTransformerBlock`].
 block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
 The tuple of output channels for each block.
 layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
 downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
 mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
 dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
 act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
 norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
 If `None`, normalization and activation layers is skipped in post-processing.
 norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
 cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
 The dimension of the cross attention features.
 transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
 The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
 [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
 [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
 reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
 The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
 blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
 [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
 [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
 encoder_hid_dim (`int`, *optional*, defaults to None):
 If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
 dimension to `cross_attention_dim`.
 encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
 If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
 embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
 attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
 num_attention_heads (`int`, *optional*):
 The number of attention heads. If not defined, defaults to `attention_head_dim`
 resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
 for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
 class_embed_type (`str`, *optional*, defaults to `None`):
 The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
 `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
 addition_embed_type (`str`, *optional*, defaults to `None`):
 Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
 "text". "text" will use the `TextTimeEmbedding` layer.
 addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
 Dimension for the timestep embeddings.
 num_class_embeds (`int`, *optional*, defaults to `None`):
 Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
 class conditioning with `class_embed_type` equal to `None`.
 time_embedding_type (`str`, *optional*, defaults to `positional`):
 The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
 time_embedding_dim (`int`, *optional*, defaults to `None`):
 An optional override for the dimension of the projected time embedding.
 time_embedding_act_fn (`str`, *optional*, defaults to `None`):
 Optional activation function to use only once on the time embeddings before they are passed to the rest of
 the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
 timestep_post_act (`str`, *optional*, defaults to `None`):
 The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
 time_cond_proj_dim (`int`, *optional*, defaults to `None`):
 The dimension of `cond_proj` layer in the timestep embedding.
 conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
 conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
 projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
 `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
 class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
 embeddings with the class embeddings.
 mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
 Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
 `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
 `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
 otherwise.
 """

_supports_gradient_checkpointing=True
_no_split_modules= ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D"]
_skip_layerwise_casting_patterns= ["norm"]

@register_to_config
def__init__(
self,
sample_size: Optional[Union[int, Tuple[int, int]]] =None,
in_channels: int=4,
out_channels: int=4,
center_input_sample: bool=False,
flip_sin_to_cos: bool=True,
freq_shift: int=0,
down_block_types: Tuple[str] = (
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"DownBlock2D",
 ),
mid_block_type: Optional[str] ="UNetMidBlock2DCrossAttn",
up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
only_cross_attention: Union[bool, Tuple[bool]] =False,
block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
layers_per_block: Union[int, Tuple[int]] =2,
downsample_padding: int=1,
mid_block_scale_factor: float=1,
dropout: float=0.0,
act_fn: str="silu",
norm_num_groups: Optional[int] =32,
norm_eps: float=1e-5,
cross_attention_dim: Union[int, Tuple[int]] =1280,
transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] =1,
reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] =None,
encoder_hid_dim: Optional[int] =None,
encoder_hid_dim_type: Optional[str] =None,
attention_head_dim: Union[int, Tuple[int]] =8,
num_attention_heads: Optional[Union[int, Tuple[int]]] =None,
dual_cross_attention: bool=False,
use_linear_projection: bool=False,
class_embed_type: Optional[str] =None,
addition_embed_type: Optional[str] =None,
addition_time_embed_dim: Optional[int] =None,
num_class_embeds: Optional[int] =None,
upcast_attention: bool=False,
resnet_time_scale_shift: str="default",
resnet_skip_time_act: bool=False,
resnet_out_scale_factor: float=1.0,
time_embedding_type: str="positional",
time_embedding_dim: Optional[int] =None,
time_embedding_act_fn: Optional[str] =None,
timestep_post_act: Optional[str] =None,
time_cond_proj_dim: Optional[int] =None,
conv_in_kernel: int=3,
conv_out_kernel: int=3,
projection_class_embeddings_input_dim: Optional[int] =None,
attention_type: str="default",
class_embeddings_concat: bool=False,
mid_block_only_cross_attention: Optional[bool] =None,
cross_attention_norm: Optional[str] =None,
addition_embed_type_num_heads: int=64,
 ):
super().__init__()

self.sample_size=sample_size

ifnum_attention_headsisnotNone:
raiseValueError(
"At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
 )

# If `num_attention_heads` is not defined (which is the case for most models)
# it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
# The reason for this behavior is to correct for incorrectly named variables that were introduced
# when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
# Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
# which is why we correct for the naming here.
num_attention_heads=num_attention_headsorattention_head_dim

# Check inputs
self._check_config(
down_block_types=down_block_types,
up_block_types=up_block_types,
only_cross_attention=only_cross_attention,
block_out_channels=block_out_channels,
layers_per_block=layers_per_block,
cross_attention_dim=cross_attention_dim,
transformer_layers_per_block=transformer_layers_per_block,
reverse_transformer_layers_per_block=reverse_transformer_layers_per_block,
attention_head_dim=attention_head_dim,
num_attention_heads=num_attention_heads,
 )

# input
conv_in_padding= (conv_in_kernel-1) //2
self.conv_in=nn.Conv2d(
in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
 )

# time
time_embed_dim, timestep_input_dim=self._set_time_proj(
time_embedding_type,
block_out_channels=block_out_channels,
flip_sin_to_cos=flip_sin_to_cos,
freq_shift=freq_shift,
time_embedding_dim=time_embedding_dim,
 )

self.time_embedding=TimestepEmbedding(
timestep_input_dim,
time_embed_dim,
act_fn=act_fn,
post_act_fn=timestep_post_act,
cond_proj_dim=time_cond_proj_dim,
 )

self._set_encoder_hid_proj(
encoder_hid_dim_type,
cross_attention_dim=cross_attention_dim,
encoder_hid_dim=encoder_hid_dim,
 )

# class embedding
self._set_class_embedding(
class_embed_type,
act_fn=act_fn,
num_class_embeds=num_class_embeds,
projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
time_embed_dim=time_embed_dim,
timestep_input_dim=timestep_input_dim,
 )

self._set_add_embedding(
addition_embed_type,
addition_embed_type_num_heads=addition_embed_type_num_heads,
addition_time_embed_dim=addition_time_embed_dim,
cross_attention_dim=cross_attention_dim,
encoder_hid_dim=encoder_hid_dim,
flip_sin_to_cos=flip_sin_to_cos,
freq_shift=freq_shift,
projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
time_embed_dim=time_embed_dim,
 )

iftime_embedding_act_fnisNone:
self.time_embed_act=None
else:
self.time_embed_act=get_activation(time_embedding_act_fn)

self.down_blocks=nn.ModuleList([])
self.up_blocks=nn.ModuleList([])

ifisinstance(only_cross_attention, bool):
ifmid_block_only_cross_attentionisNone:
mid_block_only_cross_attention=only_cross_attention

only_cross_attention= [only_cross_attention] *len(down_block_types)

ifmid_block_only_cross_attentionisNone:
mid_block_only_cross_attention=False

ifisinstance(num_attention_heads, int):
num_attention_heads= (num_attention_heads,) *len(down_block_types)

ifisinstance(attention_head_dim, int):
attention_head_dim= (attention_head_dim,) *len(down_block_types)

ifisinstance(cross_attention_dim, int):
cross_attention_dim= (cross_attention_dim,) *len(down_block_types)

ifisinstance(layers_per_block, int):
layers_per_block= [layers_per_block] *len(down_block_types)

ifisinstance(transformer_layers_per_block, int):
transformer_layers_per_block= [transformer_layers_per_block] *len(down_block_types)

ifclass_embeddings_concat:
# The time embeddings are concatenated with the class embeddings. The dimension of the
# time embeddings passed to the down, middle, and up blocks is twice the dimension of the
# regular time embeddings
blocks_time_embed_dim=time_embed_dim*2
else:
blocks_time_embed_dim=time_embed_dim

# down
output_channel=block_out_channels[0]
fori, down_block_typeinenumerate(down_block_types):
input_channel=output_channel
output_channel=block_out_channels[i]
is_final_block=i==len(block_out_channels) -1

down_block=get_down_block(
down_block_type,
num_layers=layers_per_block[i],
transformer_layers_per_block=transformer_layers_per_block[i],
in_channels=input_channel,
out_channels=output_channel,
temb_channels=blocks_time_embed_dim,
add_downsample=notis_final_block,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
cross_attention_dim=cross_attention_dim[i],
num_attention_heads=num_attention_heads[i],
downsample_padding=downsample_padding,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention[i],
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
resnet_skip_time_act=resnet_skip_time_act,
resnet_out_scale_factor=resnet_out_scale_factor,
cross_attention_norm=cross_attention_norm,
attention_head_dim=attention_head_dim[i] ifattention_head_dim[i] isnotNoneelseoutput_channel,
dropout=dropout,
 )
self.down_blocks.append(down_block)

# mid
self.mid_block=get_mid_block(
mid_block_type,
temb_channels=blocks_time_embed_dim,
in_channels=block_out_channels[-1],
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
output_scale_factor=mid_block_scale_factor,
transformer_layers_per_block=transformer_layers_per_block[-1],
num_attention_heads=num_attention_heads[-1],
cross_attention_dim=cross_attention_dim[-1],
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
mid_block_only_cross_attention=mid_block_only_cross_attention,
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
resnet_skip_time_act=resnet_skip_time_act,
cross_attention_norm=cross_attention_norm,
attention_head_dim=attention_head_dim[-1],
dropout=dropout,
 )

# count how many layers upsample the images
self.num_upsamplers=0

# up
reversed_block_out_channels=list(reversed(block_out_channels))
reversed_num_attention_heads=list(reversed(num_attention_heads))
reversed_layers_per_block=list(reversed(layers_per_block))
reversed_cross_attention_dim=list(reversed(cross_attention_dim))
reversed_transformer_layers_per_block= (
list(reversed(transformer_layers_per_block))
ifreverse_transformer_layers_per_blockisNone
elsereverse_transformer_layers_per_block
 )
only_cross_attention=list(reversed(only_cross_attention))

output_channel=reversed_block_out_channels[0]
fori, up_block_typeinenumerate(up_block_types):
is_final_block=i==len(block_out_channels) -1

prev_output_channel=output_channel
output_channel=reversed_block_out_channels[i]
input_channel=reversed_block_out_channels[min(i+1, len(block_out_channels) -1)]

# add upsample block for all BUT final layer
ifnotis_final_block:
add_upsample=True
self.num_upsamplers+=1
else:
add_upsample=False

up_block=get_up_block(
up_block_type,
num_layers=reversed_layers_per_block[i] +1,
transformer_layers_per_block=reversed_transformer_layers_per_block[i],
in_channels=input_channel,
out_channels=output_channel,
prev_output_channel=prev_output_channel,
temb_channels=blocks_time_embed_dim,
add_upsample=add_upsample,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
resolution_idx=i,
resnet_groups=norm_num_groups,
cross_attention_dim=reversed_cross_attention_dim[i],
num_attention_heads=reversed_num_attention_heads[i],
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention[i],
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
resnet_skip_time_act=resnet_skip_time_act,
resnet_out_scale_factor=resnet_out_scale_factor,
cross_attention_norm=cross_attention_norm,
attention_head_dim=attention_head_dim[i] ifattention_head_dim[i] isnotNoneelseoutput_channel,
dropout=dropout,
 )
self.up_blocks.append(up_block)

# out
ifnorm_num_groupsisnotNone:
self.conv_norm_out=nn.GroupNorm(
num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
 )

self.conv_act=get_activation(act_fn)

else:
self.conv_norm_out=None
self.conv_act=None

conv_out_padding= (conv_out_kernel-1) //2
self.conv_out=nn.Conv2d(
block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
 )

self._set_pos_net_if_use_gligen(attention_type=attention_type, cross_attention_dim=cross_attention_dim)

def_check_config(
self,
down_block_types: Tuple[str],
up_block_types: Tuple[str],
only_cross_attention: Union[bool, Tuple[bool]],
block_out_channels: Tuple[int],
layers_per_block: Union[int, Tuple[int]],
cross_attention_dim: Union[int, Tuple[int]],
transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],
reverse_transformer_layers_per_block: bool,
attention_head_dim: int,
num_attention_heads: Optional[Union[int, Tuple[int]]],
 ):
iflen(down_block_types) !=len(up_block_types):
raiseValueError(
f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
 )

iflen(block_out_channels) !=len(down_block_types):
raiseValueError(
f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
 )

ifnotisinstance(only_cross_attention, bool) andlen(only_cross_attention) !=len(down_block_types):
raiseValueError(
f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
 )

ifnotisinstance(num_attention_heads, int) andlen(num_attention_heads) !=len(down_block_types):
raiseValueError(
f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
 )

ifnotisinstance(attention_head_dim, int) andlen(attention_head_dim) !=len(down_block_types):
raiseValueError(
f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
 )

ifisinstance(cross_attention_dim, list) andlen(cross_attention_dim) !=len(down_block_types):
raiseValueError(
f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
 )

ifnotisinstance(layers_per_block, int) andlen(layers_per_block) !=len(down_block_types):
raiseValueError(
f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
 )
ifisinstance(transformer_layers_per_block, list) andreverse_transformer_layers_per_blockisNone:
forlayer_number_per_blockintransformer_layers_per_block:
ifisinstance(layer_number_per_block, list):
raiseValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")

def_set_time_proj(
self,
time_embedding_type: str,
block_out_channels: int,
flip_sin_to_cos: bool,
freq_shift: float,
time_embedding_dim: int,
 ) ->Tuple[int, int]:
iftime_embedding_type=="fourier":
time_embed_dim=time_embedding_dimorblock_out_channels[0] *2
iftime_embed_dim%2!=0:
raiseValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
self.time_proj=GaussianFourierProjection(
time_embed_dim//2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
 )
timestep_input_dim=time_embed_dim
eliftime_embedding_type=="positional":
time_embed_dim=time_embedding_dimorblock_out_channels[0] *4

self.time_proj=Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
timestep_input_dim=block_out_channels[0]
else:
raiseValueError(
f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
 )

returntime_embed_dim, timestep_input_dim

def_set_encoder_hid_proj(
self,
encoder_hid_dim_type: Optional[str],
cross_attention_dim: Union[int, Tuple[int]],
encoder_hid_dim: Optional[int],
 ):
ifencoder_hid_dim_typeisNoneandencoder_hid_dimisnotNone:
encoder_hid_dim_type="text_proj"
self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")

ifencoder_hid_dimisNoneandencoder_hid_dim_typeisnotNone:
raiseValueError(
f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
 )

ifencoder_hid_dim_type=="text_proj":
self.encoder_hid_proj=nn.Linear(encoder_hid_dim, cross_attention_dim)
elifencoder_hid_dim_type=="text_image_proj":
# image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
# case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)`
self.encoder_hid_proj=TextImageProjection(
text_embed_dim=encoder_hid_dim,
image_embed_dim=cross_attention_dim,
cross_attention_dim=cross_attention_dim,
 )
elifencoder_hid_dim_type=="image_proj":
# Kandinsky 2.2
self.encoder_hid_proj=ImageProjection(
image_embed_dim=encoder_hid_dim,
cross_attention_dim=cross_attention_dim,
 )
elifencoder_hid_dim_typeisnotNone:
raiseValueError(
f"`encoder_hid_dim_type`: {encoder_hid_dim_type} must be None, 'text_proj', 'text_image_proj', or 'image_proj'."
 )
else:
self.encoder_hid_proj=None

def_set_class_embedding(
self,
class_embed_type: Optional[str],
act_fn: str,
num_class_embeds: Optional[int],
projection_class_embeddings_input_dim: Optional[int],
time_embed_dim: int,
timestep_input_dim: int,
 ):
ifclass_embed_typeisNoneandnum_class_embedsisnotNone:
self.class_embedding=nn.Embedding(num_class_embeds, time_embed_dim)
elifclass_embed_type=="timestep":
self.class_embedding=TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
elifclass_embed_type=="identity":
self.class_embedding=nn.Identity(time_embed_dim, time_embed_dim)
elifclass_embed_type=="projection":
ifprojection_class_embeddings_input_dimisNone:
raiseValueError(
"`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
 )
# The projection `class_embed_type` is the same as the timestep `class_embed_type` except
# 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
# 2. it projects from an arbitrary input dimension.
#
# Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
# When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
# As a result, `TimestepEmbedding` can be passed arbitrary vectors.
self.class_embedding=TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
elifclass_embed_type=="simple_projection":
ifprojection_class_embeddings_input_dimisNone:
raiseValueError(
"`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
 )
self.class_embedding=nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
else:
self.class_embedding=None

def_set_add_embedding(
self,
addition_embed_type: str,
addition_embed_type_num_heads: int,
addition_time_embed_dim: Optional[int],
flip_sin_to_cos: bool,
freq_shift: float,
cross_attention_dim: Optional[int],
encoder_hid_dim: Optional[int],
projection_class_embeddings_input_dim: Optional[int],
time_embed_dim: int,
 ):
ifaddition_embed_type=="text":
ifencoder_hid_dimisnotNone:
text_time_embedding_from_dim=encoder_hid_dim
else:
text_time_embedding_from_dim=cross_attention_dim

self.add_embedding=TextTimeEmbedding(
text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
 )
elifaddition_embed_type=="text_image":
# text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
# case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
self.add_embedding=TextImageTimeEmbedding(
text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
 )
elifaddition_embed_type=="text_time":
self.add_time_proj=Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
self.add_embedding=TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
elifaddition_embed_type=="image":
# Kandinsky 2.2
self.add_embedding=ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
elifaddition_embed_type=="image_hint":
# Kandinsky 2.2 ControlNet
self.add_embedding=ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
elifaddition_embed_typeisnotNone:
raiseValueError(
f"`addition_embed_type`: {addition_embed_type} must be None, 'text', 'text_image', 'text_time', 'image', or 'image_hint'."
 )

def_set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
ifattention_typein ["gated", "gated-text-image"]:
positive_len=768
ifisinstance(cross_attention_dim, int):
positive_len=cross_attention_dim
elifisinstance(cross_attention_dim, (list, tuple)):
positive_len=cross_attention_dim[0]

feature_type="text-only"ifattention_type=="gated"else"text-image"
self.position_net=GLIGENTextBoundingboxProjection(
positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
 )

@property
defattn_processors(self) ->Dict[str, AttentionProcessor]:
r"""
 Returns:
 `dict` of attention processors: A dictionary containing all attention processors used in the model with
 indexed by its weight name.
 """
# set recursively
processors= {}

deffn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
ifhasattr(module, "get_processor"):
processors[f"{name}.processor"] =module.get_processor()

forsub_name, childinmodule.named_children():
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

returnprocessors

forname, moduleinself.named_children():
fn_recursive_add_processors(name, module, processors)

returnprocessors

defset_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
r"""
 Sets the attention processor to use to compute attention.

 Parameters:
 processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
 The instantiated processor class or a dictionary of processor classes that will be set as the processor
 for **all** `Attention` layers.

 If `processor` is a dict, the key needs to define the path to the corresponding cross attention
 processor. This is strongly recommended when setting trainable attention processors.

 """
count=len(self.attn_processors.keys())

ifisinstance(processor, dict) andlen(processor) !=count:
raiseValueError(
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
 )

deffn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
ifhasattr(module, "set_processor"):
ifnotisinstance(processor, dict):
module.set_processor(processor)
else:
module.set_processor(processor.pop(f"{name}.processor"))

forsub_name, childinmodule.named_children():
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

forname, moduleinself.named_children():
fn_recursive_attn_processor(name, module, processor)

defset_default_attn_processor(self):
"""
 Disables custom attention processors and sets the default attention implementation.
 """
ifall(proc.__class__inADDED_KV_ATTENTION_PROCESSORSforprocinself.attn_processors.values()):
processor=AttnAddedKVProcessor()
elifall(proc.__class__inCROSS_ATTENTION_PROCESSORSforprocinself.attn_processors.values()):
processor=AttnProcessor()
else:
raiseValueError(
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
 )

self.set_attn_processor(processor)

defset_attention_slice(self, slice_size: Union[str, int, List[int]] ="auto"):
r"""
 Enable sliced attention computation.

 When this option is enabled, the attention module splits the input tensor in slices to compute attention in
 several steps. This is useful for saving some memory in exchange for a small decrease in speed.

 Args:
 slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
 When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
 `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
 provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
 must be a multiple of `slice_size`.
 """
sliceable_head_dims= []

deffn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
ifhasattr(module, "set_attention_slice"):
sliceable_head_dims.append(module.sliceable_head_dim)

forchildinmodule.children():
fn_recursive_retrieve_sliceable_dims(child)

# retrieve number of attention layers
formoduleinself.children():
fn_recursive_retrieve_sliceable_dims(module)

num_sliceable_layers=len(sliceable_head_dims)

ifslice_size=="auto":
# half the attention head size is usually a good trade-off between
# speed and memory
slice_size= [dim//2fordiminsliceable_head_dims]
elifslice_size=="max":
# make smallest slice possible
slice_size=num_sliceable_layers* [1]

slice_size=num_sliceable_layers* [slice_size] ifnotisinstance(slice_size, list) elseslice_size

iflen(slice_size) !=len(sliceable_head_dims):
raiseValueError(
f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
 )

foriinrange(len(slice_size)):
size=slice_size[i]
dim=sliceable_head_dims[i]
ifsizeisnotNoneandsize>dim:
raiseValueError(f"size {size} has to be smaller or equal to {dim}.")

# Recursively walk through all the children.
# Any children which exposes the set_attention_slice method
# gets the message
deffn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
ifhasattr(module, "set_attention_slice"):
module.set_attention_slice(slice_size.pop())

forchildinmodule.children():
fn_recursive_set_attention_slice(child, slice_size)

reversed_slice_size=list(reversed(slice_size))
formoduleinself.children():
fn_recursive_set_attention_slice(module, reversed_slice_size)

defenable_freeu(self, s1: float, s2: float, b1: float, b2: float):
r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.

 The suffixes after the scaling factors represent the stage blocks where they are being applied.

 Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
 are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.

 Args:
 s1 (`float`):
 Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
 mitigate the "oversmoothing effect" in the enhanced denoising process.
 s2 (`float`):
 Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
 mitigate the "oversmoothing effect" in the enhanced denoising process.
 b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
 b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
 """
fori, upsample_blockinenumerate(self.up_blocks):
setattr(upsample_block, "s1", s1)
setattr(upsample_block, "s2", s2)
setattr(upsample_block, "b1", b1)
setattr(upsample_block, "b2", b2)

defdisable_freeu(self):
"""Disables the FreeU mechanism."""
freeu_keys= {"s1", "s2", "b1", "b2"}
fori, upsample_blockinenumerate(self.up_blocks):
forkinfreeu_keys:
ifhasattr(upsample_block, k) orgetattr(upsample_block, k, None) isnotNone:
setattr(upsample_block, k, None)

deffuse_qkv_projections(self):
"""
 Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
 are fused. For cross-attention modules, key and value projection matrices are fused.

 <Tip warning={true}>

 This API is 🧪 experimental.

 </Tip>
 """
self.original_attn_processors=None

for_, attn_processorinself.attn_processors.items():
if"Added"instr(attn_processor.__class__.__name__):
raiseValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")

self.original_attn_processors=self.attn_processors

formoduleinself.modules():
ifisinstance(module, Attention):
module.fuse_projections(fuse=True)

self.set_attn_processor(FusedAttnProcessor2_0())

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

 <Tip warning={true}>

 This API is 🧪 experimental.

 </Tip>

 """
ifself.original_attn_processorsisnotNone:
self.set_attn_processor(self.original_attn_processors)

defget_time_embed(
self, sample: torch.Tensor, timestep: Union[torch.Tensor, float, int]
 ) ->Optional[torch.Tensor]:
timesteps=timestep
ifnottorch.is_tensor(timesteps):
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
# This would be a good case for the `match` statement (Python 3.10+)
is_mps=sample.device.type=="mps"
is_npu=sample.device.type=="npu"
ifisinstance(timestep, float):
dtype=torch.float32if (is_mpsoris_npu) elsetorch.float64
else:
dtype=torch.int32if (is_mpsoris_npu) elsetorch.int64
timesteps=torch.tensor([timesteps], dtype=dtype, device=sample.device)
eliflen(timesteps.shape) ==0:
timesteps=timesteps[None].to(sample.device)

# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timesteps=timesteps.expand(sample.shape[0])

t_emb=self.time_proj(timesteps)
# `Timesteps` does not contain any weights and will always return f32 tensors
# but time_embedding might actually be running in fp16. so we need to cast here.
# there might be better ways to encapsulate this.
t_emb=t_emb.to(dtype=sample.dtype)
returnt_emb

defget_class_embed(self, sample: torch.Tensor, class_labels: Optional[torch.Tensor]) ->Optional[torch.Tensor]:
class_emb=None
ifself.class_embeddingisnotNone:
ifclass_labelsisNone:
raiseValueError("class_labels should be provided when num_class_embeds > 0")

ifself.config.class_embed_type=="timestep":
class_labels=self.time_proj(class_labels)

# `Timesteps` does not contain any weights and will always return f32 tensors
# there might be better ways to encapsulate this.
class_labels=class_labels.to(dtype=sample.dtype)

class_emb=self.class_embedding(class_labels).to(dtype=sample.dtype)
returnclass_emb

defget_aug_embed(
self, emb: torch.Tensor, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
 ) ->Optional[torch.Tensor]:
aug_emb=None
ifself.config.addition_embed_type=="text":
aug_emb=self.add_embedding(encoder_hidden_states)
elifself.config.addition_embed_type=="text_image":
# Kandinsky 2.1 - style
if"image_embeds"notinadded_cond_kwargs:
raiseValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
 )

image_embs=added_cond_kwargs.get("image_embeds")
text_embs=added_cond_kwargs.get("text_embeds", encoder_hidden_states)
aug_emb=self.add_embedding(text_embs, image_embs)
elifself.config.addition_embed_type=="text_time":
# SDXL - style
if"text_embeds"notinadded_cond_kwargs:
raiseValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
 )
text_embeds=added_cond_kwargs.get("text_embeds")
if"time_ids"notinadded_cond_kwargs:
raiseValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
 )
time_ids=added_cond_kwargs.get("time_ids")
time_embeds=self.add_time_proj(time_ids.flatten())
time_embeds=time_embeds.reshape((text_embeds.shape[0], -1))
add_embeds=torch.concat([text_embeds, time_embeds], dim=-1)
add_embeds=add_embeds.to(emb.dtype)
aug_emb=self.add_embedding(add_embeds)
elifself.config.addition_embed_type=="image":
# Kandinsky 2.2 - style
if"image_embeds"notinadded_cond_kwargs:
raiseValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
 )
image_embs=added_cond_kwargs.get("image_embeds")
aug_emb=self.add_embedding(image_embs)
elifself.config.addition_embed_type=="image_hint":
# Kandinsky 2.2 ControlNet - style
if"image_embeds"notinadded_cond_kwargsor"hint"notinadded_cond_kwargs:
raiseValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
 )
image_embs=added_cond_kwargs.get("image_embeds")
hint=added_cond_kwargs.get("hint")
aug_emb=self.add_embedding(image_embs, hint)
returnaug_emb