# 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.

import inspect

import math

from importlib import import_module

from typing import Callable, List, Optional, Union

import torch

import torch.nn.functional as F

from torch import nn

from ..image_processor import IPAdapterMaskProcessor

from ..utils import deprecate, logging

from ..utils.import_utils import is_torch_npu_available, is_xformers_available

from ..utils.torch_utils import maybe_allow_in_graph

from .lora import LoRALinearLayer

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

if is_torch_npu_available():

import torch_npu

if is_xformers_available():

import xformers

import xformers.ops

else:

xformers = None

@maybe_allow_in_graph

class Attention(nn.Module):

r"""

A cross attention layer.

Parameters:

query_dim (`int`):

The number of channels in the query.

cross_attention_dim (`int`, *optional*):

The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.

heads (`int`, *optional*, defaults to 8):

The number of heads to use for multi-head attention.

dim_head (`int`, *optional*, defaults to 64):

The number of channels in each head.

dropout (`float`, *optional*, defaults to 0.0):

The dropout probability to use.

bias (`bool`, *optional*, defaults to False):

Set to `True` for the query, key, and value linear layers to contain a bias parameter.

upcast_attention (`bool`, *optional*, defaults to False):

Set to `True` to upcast the attention computation to `float32`.

upcast_softmax (`bool`, *optional*, defaults to False):

Set to `True` to upcast the softmax computation to `float32`.

cross_attention_norm (`str`, *optional*, defaults to `None`):

The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.

cross_attention_norm_num_groups (`int`, *optional*, defaults to 32):

The number of groups to use for the group norm in the cross attention.

added_kv_proj_dim (`int`, *optional*, defaults to `None`):

The number of channels to use for the added key and value projections. If `None`, no projection is used.

norm_num_groups (`int`, *optional*, defaults to `None`):

The number of groups to use for the group norm in the attention.

spatial_norm_dim (`int`, *optional*, defaults to `None`):

The number of channels to use for the spatial normalization.

out_bias (`bool`, *optional*, defaults to `True`):

Set to `True` to use a bias in the output linear layer.

scale_qk (`bool`, *optional*, defaults to `True`):

Set to `True` to scale the query and key by `1 / sqrt(dim_head)`.

only_cross_attention (`bool`, *optional*, defaults to `False`):

Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if

`added_kv_proj_dim` is not `None`.

eps (`float`, *optional*, defaults to 1e-5):

An additional value added to the denominator in group normalization that is used for numerical stability.

rescale_output_factor (`float`, *optional*, defaults to 1.0):

A factor to rescale the output by dividing it with this value.

residual_connection (`bool`, *optional*, defaults to `False`):

Set to `True` to add the residual connection to the output.

_from_deprecated_attn_block (`bool`, *optional*, defaults to `False`):

Set to `True` if the attention block is loaded from a deprecated state dict.

processor (`AttnProcessor`, *optional*, defaults to `None`):

The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and

`AttnProcessor` otherwise.

"""

def __init__(

self,

query_dim: int,

cross_attention_dim: Optional[int] = None,

heads: int = 8,

dim_head: int = 64,

dropout: float = 0.0,

bias: bool = False,

upcast_attention: bool = False,

upcast_softmax: bool = False,

cross_attention_norm: Optional[str] = None,

cross_attention_norm_num_groups: int = 32,

qk_norm: Optional[str] = None,

added_kv_proj_dim: Optional[int] = None,

norm_num_groups: Optional[int] = None,

spatial_norm_dim: Optional[int] = None,

out_bias: bool = True,

scale_qk: bool = True,

only_cross_attention: bool = False,

eps: float = 1e-5,

rescale_output_factor: float = 1.0,

residual_connection: bool = False,

_from_deprecated_attn_block: bool = False,

processor: Optional["AttnProcessor"] = None,

out_dim: int = None,

context_pre_only=None,

):

super().__init__()

self.inner_dim = out_dim if out_dim is not None else dim_head * heads

self.query_dim = query_dim

self.use_bias = bias

self.is_cross_attention = cross_attention_dim is not None

self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim

self.upcast_attention = upcast_attention

self.upcast_softmax = upcast_softmax

self.rescale_output_factor = rescale_output_factor

self.residual_connection = residual_connection

self.dropout = dropout

self.fused_projections = False

self.out_dim = out_dim if out_dim is not None else query_dim

self.context_pre_only = context_pre_only

# we make use of this private variable to know whether this class is loaded

# with an deprecated state dict so that we can convert it on the fly

self._from_deprecated_attn_block = _from_deprecated_attn_block

self.scale_qk = scale_qk

self.scale = dim_head**-0.5 if self.scale_qk else 1.0

self.heads = out_dim // dim_head if out_dim is not None else heads

# for slice_size > 0 the attention score computation

# is split across the batch axis to save memory

# You can set slice_size with `set_attention_slice`

self.sliceable_head_dim = heads

self.added_kv_proj_dim = added_kv_proj_dim

self.only_cross_attention = only_cross_attention

if self.added_kv_proj_dim is None and self.only_cross_attention:

raise ValueError(

"`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."

)

if norm_num_groups is not None:

self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True)

else:

self.group_norm = None

if spatial_norm_dim is not None:

self.spatial_norm = SpatialNorm(f_channels=query_dim, zq_channels=spatial_norm_dim)

else:

self.spatial_norm = None

if qk_norm is None:

self.norm_q = None

self.norm_k = None

elif qk_norm == "layer_norm":

self.norm_q = nn.LayerNorm(dim_head, eps=eps)

self.norm_k = nn.LayerNorm(dim_head, eps=eps)

else:

raise ValueError(f"unknown qk_norm: {qk_norm}. Should be None or 'layer_norm'")

if cross_attention_norm is None:

self.norm_cross = None

elif cross_attention_norm == "layer_norm":

self.norm_cross = nn.LayerNorm(self.cross_attention_dim)

elif cross_attention_norm == "group_norm":

if self.added_kv_proj_dim is not None:

# The given `encoder_hidden_states` are initially of shape

# (batch_size, seq_len, added_kv_proj_dim) before being projected

# to (batch_size, seq_len, cross_attention_dim). The norm is applied

# before the projection, so we need to use `added_kv_proj_dim` as

# the number of channels for the group norm.

norm_cross_num_channels = added_kv_proj_dim

else:

norm_cross_num_channels = self.cross_attention_dim

self.norm_cross = nn.GroupNorm(

num_channels=norm_cross_num_channels, num_groups=cross_attention_norm_num_groups, eps=1e-5, affine=True

)

else:

raise ValueError(

f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'"

)

self.to_q = nn.Linear(query_dim, self.inner_dim, bias=bias)

if not self.only_cross_attention:

# only relevant for the `AddedKVProcessor` classes

self.to_k = nn.Linear(self.cross_attention_dim, self.inner_dim, bias=bias)

self.to_v = nn.Linear(self.cross_attention_dim, self.inner_dim, bias=bias)

else:

self.to_k = None

self.to_v = None

if self.added_kv_proj_dim is not None:

self.add_k_proj = nn.Linear(added_kv_proj_dim, self.inner_dim)

self.add_v_proj = nn.Linear(added_kv_proj_dim, self.inner_dim)

if self.context_pre_only is not None:

self.add_q_proj = nn.Linear(added_kv_proj_dim, self.inner_dim)

self.to_out = nn.ModuleList([])

self.to_out.append(nn.Linear(self.inner_dim, self.out_dim, bias=out_bias))

self.to_out.append(nn.Dropout(dropout))

if self.context_pre_only is not None and not self.context_pre_only:

self.to_add_out = nn.Linear(self.inner_dim, self.out_dim, bias=out_bias)

# set attention processor

# We use the AttnProcessor2_0 by default when torch 2.x is used which uses

# torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention

# but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1

if processor is None:

processor = (

AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()

)

self.set_processor(processor)

def set_use_npu_flash_attention(self, use_npu_flash_attention: bool) -> None:

r"""

Set whether to use npu flash attention from `torch_npu` or not.

"""

if use_npu_flash_attention:

processor = AttnProcessorNPU()

else:

# set attention processor

# We use the AttnProcessor2_0 by default when torch 2.x is used which uses

# torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention

# but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1

processor = (

AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()

)

self.set_processor(processor)

def set_use_memory_efficient_attention_xformers(

self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None

) -> None:

r"""

Set whether to use memory efficient attention from `xformers` or not.

Args:

use_memory_efficient_attention_xformers (`bool`):

Whether to use memory efficient attention from `xformers` or not.

attention_op (`Callable`, *optional*):

The attention operation to use. Defaults to `None` which uses the default attention operation from

`xformers`.

"""

is_lora = hasattr(self, "processor") and isinstance(

self.processor,

LORA_ATTENTION_PROCESSORS,

)

is_custom_diffusion = hasattr(self, "processor") and isinstance(

self.processor,

(CustomDiffusionAttnProcessor, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0),

)

is_added_kv_processor = hasattr(self, "processor") and isinstance(

self.processor,

(

AttnAddedKVProcessor,

AttnAddedKVProcessor2_0,

SlicedAttnAddedKVProcessor,

XFormersAttnAddedKVProcessor,

LoRAAttnAddedKVProcessor,

),

)

if use_memory_efficient_attention_xformers:

if is_added_kv_processor and (is_lora or is_custom_diffusion):

raise NotImplementedError(

f"Memory efficient attention is currently not supported for LoRA or custom diffusion for attention processor type {self.processor}"

)

if not is_xformers_available():

raise ModuleNotFoundError(

(

"Refer to https://github.com/facebookresearch/xformers for more information on how to install"

" xformers"

),

name="xformers",

)

elif not torch.cuda.is_available():

raise ValueError(

"torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"

" only available for GPU "

)

else:

try:

# Make sure we can run the memory efficient attention

_ = xformers.ops.memory_efficient_attention(

torch.randn((1, 2, 40), device="cuda"),

torch.randn((1, 2, 40), device="cuda"),

torch.randn((1, 2, 40), device="cuda"),

)

except Exception as e:

raise e

if is_lora:

# TODO (sayakpaul): should we throw a warning if someone wants to use the xformers

# variant when using PT 2.0 now that we have LoRAAttnProcessor2_0?

processor = LoRAXFormersAttnProcessor(

hidden_size=self.processor.hidden_size,

cross_attention_dim=self.processor.cross_attention_dim,

rank=self.processor.rank,

attention_op=attention_op,

)

processor.load_state_dict(self.processor.state_dict())

processor.to(self.processor.to_q_lora.up.weight.device)

elif is_custom_diffusion:

processor = CustomDiffusionXFormersAttnProcessor(

train_kv=self.processor.train_kv,

train_q_out=self.processor.train_q_out,

hidden_size=self.processor.hidden_size,

cross_attention_dim=self.processor.cross_attention_dim,

attention_op=attention_op,

)

processor.load_state_dict(self.processor.state_dict())

if hasattr(self.processor, "to_k_custom_diffusion"):

processor.to(self.processor.to_k_custom_diffusion.weight.device)

elif is_added_kv_processor:

# TODO(Patrick, Suraj, William) - currently xformers doesn't work for UnCLIP

# which uses this type of cross attention ONLY because the attention mask of format

# [0, ..., -10.000, ..., 0, ...,] is not supported

# throw warning

logger.info(

"Memory efficient attention with `xformers` might currently not work correctly if an attention mask is required for the attention operation."

)

processor = XFormersAttnAddedKVProcessor(attention_op=attention_op)

else:

processor = XFormersAttnProcessor(attention_op=attention_op)

else:

if is_lora:

attn_processor_class = (

LoRAAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else LoRAAttnProcessor

)

processor = attn_processor_class(

hidden_size=self.processor.hidden_size,

cross_attention_dim=self.processor.cross_attention_dim,

rank=self.processor.rank,

)

processor.load_state_dict(self.processor.state_dict())

processor.to(self.processor.to_q_lora.up.weight.device)

elif is_custom_diffusion:

attn_processor_class = (

CustomDiffusionAttnProcessor2_0

if hasattr(F, "scaled_dot_product_attention")

else CustomDiffusionAttnProcessor

)

processor = attn_processor_class(

train_kv=self.processor.train_kv,

train_q_out=self.processor.train_q_out,

hidden_size=self.processor.hidden_size,

cross_attention_dim=self.processor.cross_attention_dim,

)

processor.load_state_dict(self.processor.state_dict())

if hasattr(self.processor, "to_k_custom_diffusion"):

processor.to(self.processor.to_k_custom_diffusion.weight.device)

else:

# set attention processor

# We use the AttnProcessor2_0 by default when torch 2.x is used which uses

# torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention

# but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1

processor = (

AttnProcessor2_0()

if hasattr(F, "scaled_dot_product_attention") and self.scale_qk

else AttnProcessor()

)

self.set_processor(processor)

def set_attention_slice(self, slice_size: int) -> None:

r"""

Set the slice size for attention computation.

Args:

slice_size (`int`):

The slice size for attention computation.

"""

if slice_size is not None and slice_size > self.sliceable_head_dim:

raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")

if slice_size is not None and self.added_kv_proj_dim is not None:

processor = SlicedAttnAddedKVProcessor(slice_size)

elif slice_size is not None:

processor = SlicedAttnProcessor(slice_size)

elif self.added_kv_proj_dim is not None:

processor = AttnAddedKVProcessor()

else:

# set attention processor

# We use the AttnProcessor2_0 by default when torch 2.x is used which uses

# torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention

# but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1

processor = (

AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()

)

self.set_processor(processor)

def set_processor(self, processor: "AttnProcessor") -> None:

r"""

Set the attention processor to use.

Args:

processor (`AttnProcessor`):

The attention processor to use.

"""

# if current processor is in `self._modules` and if passed `processor` is not, we need to

# pop `processor` from `self._modules`

if (

hasattr(self, "processor")

and isinstance(self.processor, torch.nn.Module)

and not isinstance(processor, torch.nn.Module)

):

logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}")

self._modules.pop("processor")

self.processor = processor

def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor":

r"""

Get the attention processor in use.

Args:

return_deprecated_lora (`bool`, *optional*, defaults to `False`):

Set to `True` to return the deprecated LoRA attention processor.

Returns:

"AttentionProcessor": The attention processor in use.

"""

if not return_deprecated_lora:

return self.processor

# TODO(Sayak, Patrick). The rest of the function is needed to ensure backwards compatible

# serialization format for LoRA Attention Processors. It should be deleted once the integration

# with PEFT is completed.

is_lora_activated = {

name: module.lora_layer is not None

for name, module in self.named_modules()

if hasattr(module, "lora_layer")

}

# 1. if no layer has a LoRA activated we can return the processor as usual

if not any(is_lora_activated.values()):

return self.processor

# If doesn't apply LoRA do `add_k_proj` or `add_v_proj`

is_lora_activated.pop("add_k_proj", None)

is_lora_activated.pop("add_v_proj", None)

# 2. else it is not possible that only some layers have LoRA activated

if not all(is_lora_activated.values()):

raise ValueError(

f"Make sure that either all layers or no layers have LoRA activated, but have {is_lora_activated}"

)

# 3. And we need to merge the current LoRA layers into the corresponding LoRA attention processor

non_lora_processor_cls_name = self.processor.__class__.__name__

lora_processor_cls = getattr(import_module(__name__), "LoRA" + non_lora_processor_cls_name)

hidden_size = self.inner_dim

# now create a LoRA attention processor from the LoRA layers

if lora_processor_cls in [LoRAAttnProcessor, LoRAAttnProcessor2_0, LoRAXFormersAttnProcessor]:

kwargs = {

"cross_attention_dim": self.cross_attention_dim,

"rank": self.to_q.lora_layer.rank,

"network_alpha": self.to_q.lora_layer.network_alpha,

"q_rank": self.to_q.lora_layer.rank,

"q_hidden_size": self.to_q.lora_layer.out_features,

"k_rank": self.to_k.lora_layer.rank,

"k_hidden_size": self.to_k.lora_layer.out_features,

"v_rank": self.to_v.lora_layer.rank,

"v_hidden_size": self.to_v.lora_layer.out_features,

"out_rank": self.to_out[0].lora_layer.rank,

"out_hidden_size": self.to_out[0].lora_layer.out_features,

}

if hasattr(self.processor, "attention_op"):

kwargs["attention_op"] = self.processor.attention_op

lora_processor = lora_processor_cls(hidden_size, **kwargs)

lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict())

lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict())

lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict())

lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict())

elif lora_processor_cls == LoRAAttnAddedKVProcessor:

lora_processor = lora_processor_cls(

hidden_size,

cross_attention_dim=self.add_k_proj.weight.shape[0],

rank=self.to_q.lora_layer.rank,

network_alpha=self.to_q.lora_layer.network_alpha,

)

lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict())

lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict())

lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict())

lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict())

# only save if used

if self.add_k_proj.lora_layer is not None:

lora_processor.add_k_proj_lora.load_state_dict(self.add_k_proj.lora_layer.state_dict())

lora_processor.add_v_proj_lora.load_state_dict(self.add_v_proj.lora_layer.state_dict())

else:

lora_processor.add_k_proj_lora = None

lora_processor.add_v_proj_lora = None

else:

raise ValueError(f"{lora_processor_cls} does not exist.")

return lora_processor

def forward(

self,

hidden_states: torch.Tensor,

encoder_hidden_states: Optional[torch.Tensor] = None,

attention_mask: Optional[torch.Tensor] = None,

**cross_attention_kwargs,

) -> torch.Tensor:

r"""

The forward method of the `Attention` class.

Args:

hidden_states (`torch.Tensor`):

The hidden states of the query.

encoder_hidden_states (`torch.Tensor`, *optional*):

The hidden states of the encoder.

attention_mask (`torch.Tensor`, *optional*):

The attention mask to use. If `None`, no mask is applied.

**cross_attention_kwargs:

Additional keyword arguments to pass along to the cross attention.

Returns:

`torch.Tensor`: The output of the attention layer.

"""

# The `Attention` class can call different attention processors / attention functions

# here we simply pass along all tensors to the selected processor class

# For standard processors that are defined here, `**cross_attention_kwargs` is empty

attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())

quiet_attn_parameters = {"ip_adapter_masks"}

unused_kwargs = [

k for k, _ in cross_attention_kwargs.items() if k not in attn_parameters and k not in quiet_attn_parameters

]

if len(unused_kwargs) > 0:

logger.warning(

f"cross_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored."

)

cross_attention_kwargs = {k: w for k, w in cross_attention_kwargs.items() if k in attn_parameters}

return self.processor(

self,

hidden_states,

encoder_hidden_states=encoder_hidden_states,

attention_mask=attention_mask,

**cross_attention_kwargs,

)

def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor:

r"""

Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`. `heads`

is the number of heads initialized while constructing the `Attention` class.

Args:

tensor (`torch.Tensor`): The tensor to reshape.

Returns:

`torch.Tensor`: The reshaped tensor.

"""

head_size = self.heads

batch_size, seq_len, dim = tensor.shape

tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)

tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)

return tensor

def head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor:

r"""

Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size, seq_len, heads, dim // heads]` `heads` is

the number of heads initialized while constructing the `Attention` class.

Args:

tensor (`torch.Tensor`): The tensor to reshape.

out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor. If `3`, the tensor is

reshaped to `[batch_size * heads, seq_len, dim // heads]`.

Returns:

`torch.Tensor`: The reshaped tensor.

"""

head_size = self.heads

if tensor.ndim == 3:

batch_size, seq_len, dim = tensor.shape

extra_dim = 1

else:

batch_size, extra_dim, seq_len, dim = tensor.shape

tensor = tensor.reshape(batch_size, seq_len * extra_dim, head_size, dim // head_size)

tensor = tensor.permute(0, 2, 1, 3)

if out_dim == 3:

tensor = tensor.reshape(batch_size * head_size, seq_len * extra_dim, dim // head_size)

return tensor

def get_attention_scores(

self, query: torch.Tensor, key: torch.Tensor, attention_mask: torch.Tensor = None

) -> torch.Tensor:

r"""

Compute the attention scores.

Args:

query (`torch.Tensor`): The query tensor.

key (`torch.Tensor`): The key tensor.

attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied.

Returns:

`torch.Tensor`: The attention probabilities/scores.

"""

dtype = query.dtype

if self.upcast_attention:

query = query.float()

key = key.float()

if attention_mask is None:

baddbmm_input = torch.empty(

query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device

)

beta = 0

else:

baddbmm_input = attention_mask

beta = 1

attention_scores = torch.baddbmm(

baddbmm_input,

query,

key.transpose(-1, -2),

beta=beta,

alpha=self.scale,

)

del baddbmm_input

if self.upcast_softmax:

attention_scores = attention_scores.float()

attention_probs = attention_scores.softmax(dim=-1)

del attention_scores

attention_probs = attention_probs.to(dtype)

return attention_probs

def prepare_attention_mask(

self, attention_mask: torch.Tensor, target_length: int, batch_size: int, out_dim: int = 3

) -> torch.Tensor:

r"""

Prepare the attention mask for the attention computation.

Args:

attention_mask (`torch.Tensor`):

The attention mask to prepare.

target_length (`int`):

The target length of the attention mask. This is the length of the attention mask after padding.

batch_size (`int`):

The batch size, which is used to repeat the attention mask.

out_dim (`int`, *optional*, defaults to `3`):

The output dimension of the attention mask. Can be either `3` or `4`.

Returns:

`torch.Tensor`: The prepared attention mask.

"""

head_size = self.heads

if attention_mask is None:

return attention_mask

current_length: int = attention_mask.shape[-1]

if current_length != target_length:

if attention_mask.device.type == "mps":

# HACK: MPS: Does not support padding by greater than dimension of input tensor.

# Instead, we can manually construct the padding tensor.

padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length)

padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device)

attention_mask = torch.cat([attention_mask, padding], dim=2)

else:

# TODO: for pipelines such as stable-diffusion, padding cross-attn mask:

# we want to instead pad by (0, remaining_length), where remaining_length is:

# remaining_length: int = target_length - current_length

# TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding

attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)

if out_dim == 3:

if attention_mask.shape[0] < batch_size * head_size:

attention_mask = attention_mask.repeat_interleave(head_size, dim=0)

elif out_dim == 4:

attention_mask = attention_mask.unsqueeze(1)

attention_mask = attention_mask.repeat_interleave(head_size, dim=1)

return attention_mask

def norm_encoder_hidden_states(self, encoder_hidden_states: torch.Tensor) -> torch.Tensor:

r"""

Normalize the encoder hidden states. Requires `self.norm_cross` to be specified when constructing the

`Attention` class.

Args:

encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder.

Returns:

`torch.Tensor`: The normalized encoder hidden states.

"""

assert self.norm_cross is not None, "self.norm_cross must be defined to call self.norm_encoder_hidden_states"

if isinstance(self.norm_cross, nn.LayerNorm):

encoder_hidden_states = self.norm_cross(encoder_hidden_states)

elif isinstance(self.norm_cross, nn.GroupNorm):

# Group norm norms along the channels dimension and expects

# input to be in the shape of (N, C, *). In this case, we want

# to norm along the hidden dimension, so we need to move

# (batch_size, sequence_length, hidden_size) ->

# (batch_size, hidden_size, sequence_length)

encoder_hidden_states = encoder_hidden_states.transpose(1, 2)

encoder_hidden_states = self.norm_cross(encoder_hidden_states)

encoder_hidden_states = encoder_hidden_states.transpose(1, 2)

else:

assert False

return encoder_hidden_states

@torch.no_grad()

def fuse_projections(self, fuse=True):

device = self.to_q.weight.data.device

dtype = self.to_q.weight.data.dtype

if not self.is_cross_attention:

# fetch weight matrices.

concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data])

in_features = concatenated_weights.shape[1]

out_features = concatenated_weights.shape[0]

# create a new single projection layer and copy over the weights.

self.to_qkv = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)

self.to_qkv.weight.copy_(concatenated_weights)

if self.use_bias:

concatenated_bias = torch.cat([self.to_q.bias.data, self.to_k.bias.data, self.to_v.bias.data])

self.to_qkv.bias.copy_(concatenated_bias)

else:

concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data])

in_features = concatenated_weights.shape[1]

out_features = concatenated_weights.shape[0]

self.to_kv = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)

self.to_kv.weight.copy_(concatenated_weights)

if self.use_bias:

concatenated_bias = torch.cat([self.to_k.bias.data, self.to_v.bias.data])

self.to_kv.bias.copy_(concatenated_bias)

self.fused_projections = fuse

class AttnProcessor:

r"""

Default processor for performing attention-related computations.

"""

def __call__(

self,

attn: Attention,

hidden_states: torch.Tensor,

encoder_hidden_states: Optional[torch.Tensor] = None,

attention_mask: Optional[torch.Tensor] = None,

temb: Optional[torch.Tensor] = None,

*args,

**kwargs,

) -> torch.Tensor:

if len(args) > 0 or kwargs.get("scale", None) is not None:

deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."

deprecate("scale", "1.0.0", deprecation_message)

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)

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)

# 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 CustomDiffusionAttnProcessor(nn.Module):

r"""

Processor for implementing attention for the Custom Diffusion method.

Args:

train_kv (`bool`, defaults to `True`):

Whether to newly train the key and value matrices corresponding to the text features.

train_q_out (`bool`, defaults to `True`):

Whether to newly train query matrices corresponding to the latent image features.

hidden_size (`int`, *optional*, defaults to `None`):

The hidden size of the attention layer.

cross_attention_dim (`int`, *optional*, defaults to `None`):

The number of channels in the `encoder_hidden_states`.

out_bias (`bool`, defaults to `True`):

Whether to include the bias parameter in `train_q_out`.

dropout (`float`, *optional*, defaults to 0.0):

The dropout probability to use.

"""

def __init__(

self,

train_kv: bool = True,

train_q_out: bool = True,

hidden_size: Optional[int] = None,

cross_attention_dim: Optional[int] = None,

out_bias: bool = True,

dropout: float = 0.0,

):

super().__init__()

self.train_kv = train_kv

self.train_q_out = train_q_out

self.hidden_size = hidden_size

self.cross_attention_dim = cross_attention_dim

# `_custom_diffusion` id for easy serialization and loading.

if self.train_kv:

self.to_k_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)

self.to_v_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)

if self.train_q_out:

self.to_q_custom_diffusion = nn.Linear(hidden_size, hidden_size, bias=False)

self.to_out_custom_diffusion = nn.ModuleList([])

self.to_out_custom_diffusion.append(nn.Linear(hidden_size, hidden_size, bias=out_bias))

self.to_out_custom_diffusion.append(nn.Dropout(dropout))

def __call__(

self,

attn: Attention,

hidden_states: torch.Tensor,

encoder_hidden_states: Optional[torch.Tensor] = None,

attention_mask: Optional[torch.Tensor] = None,

) -> torch.Tensor:

batch_size, sequence_length, _ = hidden_states.shape

attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)

if self.train_q_out:

query = self.to_q_custom_diffusion(hidden_states).to(attn.to_q.weight.dtype)

else:

query = attn.to_q(hidden_states.to(attn.to_q.weight.dtype))

if encoder_hidden_states is None:

crossattn = False

encoder_hidden_states = hidden_states

else:

crossattn = True

if attn.norm_cross:

encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

if self.train_kv:

key = self.to_k_custom_diffusion(encoder_hidden_states.to(self.to_k_custom_diffusion.weight.dtype))

value = self.to_v_custom_diffusion(encoder_hidden_states.to(self.to_v_custom_diffusion.weight.dtype))

key = key.to(attn.to_q.weight.dtype)

value = value.to(attn.to_q.weight.dtype)

else:

key = attn.to_k(encoder_hidden_states)

value = attn.to_v(encoder_hidden_states)

if crossattn:

detach = torch.ones_like(key)

detach[:, :1, :] = detach[:, :1, :] * 0.0

key = detach * key + (1 - detach) * key.detach()

value = detach * value + (1 - detach) * value.detach()

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)

if self.train_q_out:

# linear proj

hidden_states = self.to_out_custom_diffusion[0](hidden_states)

# dropout

hidden_states = self.to_out_custom_diffusion[1](hidden_states)

else:

# linear proj

hidden_states = attn.to_out[0](hidden_states)

# dropout

hidden_states = attn.to_out[1](hidden_states)

return hidden_states

class AttnAddedKVProcessor:

r"""

Processor for performing attention-related computations with extra learnable key and value matrices for the text

encoder.

"""

def __call__(

self,

attn: Attention,

hidden_states: torch.Tensor,

encoder_hidden_states: Optional[torch.Tensor] = None,

attention_mask: Optional[torch.Tensor] = None,

*args,

**kwargs,

) -> torch.Tensor:

if len(args) > 0 or kwargs.get("scale", None) is not None:

deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."

deprecate("scale", "1.0.0", deprecation_message)

residual = hidden_states

hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)

batch_size, sequence_length, _ = hidden_states.shape

attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)

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)

hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

query = attn.to_q(hidden_states)

query = attn.head_to_batch_dim(query)

encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)

encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)

encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj)

encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj)

if not attn.only_cross_attention:

key = attn.to_k(hidden_states)

value = attn.to_v(hidden_states)

key = attn.head_to_batch_dim(key)

value = attn.head_to_batch_dim(value)

key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)

value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)

else:

key = encoder_hidden_states_key_proj

value = encoder_hidden_states_value_proj

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)

# linear proj

hidden_states = attn.to_out[0](hidden_states)

# dropout