attention_processor.py 122 KB
Newer Older
1
# Copyright 2024 The HuggingFace Team. All rights reserved.
Patrick von Platen's avatar
Patrick von Platen committed
2
3
4
5
6
7
8
9
10
11
12
13
#
# 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.
14
import inspect
15
import math
16
from typing import Callable, List, Optional, Union
Patrick von Platen's avatar
Patrick von Platen committed
17
18
19

import torch
import torch.nn.functional as F
20
from torch import nn
Patrick von Platen's avatar
Patrick von Platen committed
21

22
from ..image_processor import IPAdapterMaskProcessor
23
from ..utils import deprecate, logging
24
from ..utils.import_utils import is_torch_npu_available, is_xformers_available
Dhruv Nair's avatar
Dhruv Nair committed
25
from ..utils.torch_utils import maybe_allow_in_graph
Patrick von Platen's avatar
Patrick von Platen committed
26
27
28
29


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

30
31
if is_torch_npu_available():
    import torch_npu
Patrick von Platen's avatar
Patrick von Platen committed
32
33
34
35
36
37
38
39

if is_xformers_available():
    import xformers
    import xformers.ops
else:
    xformers = None


40
@maybe_allow_in_graph
Patrick von Platen's avatar
Patrick von Platen committed
41
42
43
44
45
class Attention(nn.Module):
    r"""
    A cross attention layer.

    Parameters:
46
47
        query_dim (`int`):
            The number of channels in the query.
Patrick von Platen's avatar
Patrick von Platen committed
48
49
        cross_attention_dim (`int`, *optional*):
            The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
50
51
52
53
54
55
        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.
Patrick von Platen's avatar
Patrick von Platen committed
56
57
        bias (`bool`, *optional*, defaults to False):
            Set to `True` for the query, key, and value linear layers to contain a bias parameter.
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
        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.
Patrick von Platen's avatar
Patrick von Platen committed
90
91
92
93
94
95
96
    """

    def __init__(
        self,
        query_dim: int,
        cross_attention_dim: Optional[int] = None,
        heads: int = 8,
97
        kv_heads: Optional[int] = None,
Patrick von Platen's avatar
Patrick von Platen committed
98
99
        dim_head: int = 64,
        dropout: float = 0.0,
100
        bias: bool = False,
Patrick von Platen's avatar
Patrick von Platen committed
101
102
        upcast_attention: bool = False,
        upcast_softmax: bool = False,
103
104
        cross_attention_norm: Optional[str] = None,
        cross_attention_norm_num_groups: int = 32,
105
        qk_norm: Optional[str] = None,
Patrick von Platen's avatar
Patrick von Platen committed
106
107
        added_kv_proj_dim: Optional[int] = None,
        norm_num_groups: Optional[int] = None,
YiYi Xu's avatar
YiYi Xu committed
108
        spatial_norm_dim: Optional[int] = None,
Patrick von Platen's avatar
Patrick von Platen committed
109
110
        out_bias: bool = True,
        scale_qk: bool = True,
111
        only_cross_attention: bool = False,
112
113
114
        eps: float = 1e-5,
        rescale_output_factor: float = 1.0,
        residual_connection: bool = False,
115
        _from_deprecated_attn_block: bool = False,
Patrick von Platen's avatar
Patrick von Platen committed
116
        processor: Optional["AttnProcessor"] = None,
117
        out_dim: int = None,
Dhruv Nair's avatar
Dhruv Nair committed
118
        context_pre_only=None,
Patrick von Platen's avatar
Patrick von Platen committed
119
120
    ):
        super().__init__()
121
        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
122
        self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads
123
        self.query_dim = query_dim
124
125
        self.use_bias = bias
        self.is_cross_attention = cross_attention_dim is not None
126
        self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
Patrick von Platen's avatar
Patrick von Platen committed
127
128
        self.upcast_attention = upcast_attention
        self.upcast_softmax = upcast_softmax
129
130
        self.rescale_output_factor = rescale_output_factor
        self.residual_connection = residual_connection
131
        self.dropout = dropout
132
        self.fused_projections = False
133
        self.out_dim = out_dim if out_dim is not None else query_dim
Dhruv Nair's avatar
Dhruv Nair committed
134
        self.context_pre_only = context_pre_only
135
136
137
138

        # 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
Patrick von Platen's avatar
Patrick von Platen committed
139

140
141
        self.scale_qk = scale_qk
        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
Patrick von Platen's avatar
Patrick von Platen committed
142

143
        self.heads = out_dim // dim_head if out_dim is not None else heads
Patrick von Platen's avatar
Patrick von Platen committed
144
145
146
147
148
149
        # 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
150
151
152
153
154
155
        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`."
            )
Patrick von Platen's avatar
Patrick von Platen committed
156
157

        if norm_num_groups is not None:
158
            self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True)
Patrick von Platen's avatar
Patrick von Platen committed
159
160
161
        else:
            self.group_norm = None

YiYi Xu's avatar
YiYi Xu committed
162
163
164
165
166
        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

167
168
169
170
171
172
        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)
173
174
175
176
        elif qk_norm == "layer_norm_across_heads":
            # Lumina applys qk norm across all heads
            self.norm_q = nn.LayerNorm(dim_head * heads, eps=eps)
            self.norm_k = nn.LayerNorm(dim_head * kv_heads, eps=eps)
177
178
179
        else:
            raise ValueError(f"unknown qk_norm: {qk_norm}. Should be None or 'layer_norm'")

180
181
182
        if cross_attention_norm is None:
            self.norm_cross = None
        elif cross_attention_norm == "layer_norm":
183
            self.norm_cross = nn.LayerNorm(self.cross_attention_dim)
184
185
186
187
188
189
190
191
192
        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:
193
                norm_cross_num_channels = self.cross_attention_dim
194
195
196
197
198
199
200
201

            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'"
            )
Patrick von Platen's avatar
Patrick von Platen committed
202

203
        self.to_q = nn.Linear(query_dim, self.inner_dim, bias=bias)
204
205
206

        if not self.only_cross_attention:
            # only relevant for the `AddedKVProcessor` classes
207
208
            self.to_k = nn.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias)
            self.to_v = nn.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias)
209
210
211
        else:
            self.to_k = None
            self.to_v = None
Patrick von Platen's avatar
Patrick von Platen committed
212
213

        if self.added_kv_proj_dim is not None:
214
215
            self.add_k_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim)
            self.add_v_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim)
Dhruv Nair's avatar
Dhruv Nair committed
216
217
            if self.context_pre_only is not None:
                self.add_q_proj = nn.Linear(added_kv_proj_dim, self.inner_dim)
Patrick von Platen's avatar
Patrick von Platen committed
218
219

        self.to_out = nn.ModuleList([])
220
        self.to_out.append(nn.Linear(self.inner_dim, self.out_dim, bias=out_bias))
Patrick von Platen's avatar
Patrick von Platen committed
221
222
        self.to_out.append(nn.Dropout(dropout))

Dhruv Nair's avatar
Dhruv Nair committed
223
224
225
        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)

Patrick von Platen's avatar
Patrick von Platen committed
226
227
228
229
230
231
        # 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 = (
232
                AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
Patrick von Platen's avatar
Patrick von Platen committed
233
234
235
            )
        self.set_processor(processor)

236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
    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)

Patrick von Platen's avatar
Patrick von Platen committed
253
254
    def set_use_memory_efficient_attention_xformers(
        self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None
255
256
257
258
259
260
261
262
263
264
265
    ) -> 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`.
        """
266
        is_custom_diffusion = hasattr(self, "processor") and isinstance(
267
268
            self.processor,
            (CustomDiffusionAttnProcessor, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0),
269
        )
270
271
272
273
274
275
276
277
278
        is_added_kv_processor = hasattr(self, "processor") and isinstance(
            self.processor,
            (
                AttnAddedKVProcessor,
                AttnAddedKVProcessor2_0,
                SlicedAttnAddedKVProcessor,
                XFormersAttnAddedKVProcessor,
            ),
        )
Patrick von Platen's avatar
Patrick von Platen committed
279
280

        if use_memory_efficient_attention_xformers:
281
            if is_added_kv_processor and is_custom_diffusion:
Patrick von Platen's avatar
Patrick von Platen committed
282
                raise NotImplementedError(
283
                    f"Memory efficient attention is currently not supported for custom diffusion for attention processor type {self.processor}"
Patrick von Platen's avatar
Patrick von Platen committed
284
                )
285
            if not is_xformers_available():
Patrick von Platen's avatar
Patrick von Platen committed
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
                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

309
            if is_custom_diffusion:
310
311
312
313
314
315
316
317
318
319
                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)
320
321
322
323
324
325
326
327
328
            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)
Patrick von Platen's avatar
Patrick von Platen committed
329
330
331
            else:
                processor = XFormersAttnProcessor(attention_op=attention_op)
        else:
332
            if is_custom_diffusion:
333
334
335
336
337
338
                attn_processor_class = (
                    CustomDiffusionAttnProcessor2_0
                    if hasattr(F, "scaled_dot_product_attention")
                    else CustomDiffusionAttnProcessor
                )
                processor = attn_processor_class(
339
340
341
342
343
344
345
346
                    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)
Patrick von Platen's avatar
Patrick von Platen committed
347
            else:
348
349
350
351
352
353
354
355
356
                # 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()
                )
Patrick von Platen's avatar
Patrick von Platen committed
357
358
359

        self.set_processor(processor)

360
361
362
363
364
365
366
367
    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.
        """
Patrick von Platen's avatar
Patrick von Platen committed
368
369
370
371
372
373
374
375
376
377
        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:
378
379
380
381
382
383
384
            # 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()
            )
Patrick von Platen's avatar
Patrick von Platen committed
385
386
387

        self.set_processor(processor)

388
    def set_processor(self, processor: "AttnProcessor") -> None:
389
390
391
392
393
394
395
        r"""
        Set the attention processor to use.

        Args:
            processor (`AttnProcessor`):
                The attention processor to use.
        """
Patrick von Platen's avatar
Patrick von Platen committed
396
397
398
399
400
401
402
403
404
405
406
407
        # 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

408
    def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor":
409
410
411
412
413
414
415
416
417
418
        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.
        """
419
420
421
        if not return_deprecated_lora:
            return self.processor

422
423
    def forward(
        self,
424
425
426
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
        **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.
        """
Patrick von Platen's avatar
Patrick von Platen committed
445
446
447
        # 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
448
449

        attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
450
451
452
453
        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
        ]
454
455
456
457
458
459
        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}

Patrick von Platen's avatar
Patrick von Platen committed
460
461
462
463
464
465
466
467
        return self.processor(
            self,
            hidden_states,
            encoder_hidden_states=encoder_hidden_states,
            attention_mask=attention_mask,
            **cross_attention_kwargs,
        )

468
469
470
471
472
473
474
475
476
477
478
    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.
        """
Patrick von Platen's avatar
Patrick von Platen committed
479
480
481
482
483
484
        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

485
486
487
488
489
490
491
492
493
494
495
496
497
    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.
        """
Patrick von Platen's avatar
Patrick von Platen committed
498
        head_size = self.heads
499
500
501
502
503
504
        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)
505
506
507
        tensor = tensor.permute(0, 2, 1, 3)

        if out_dim == 3:
508
            tensor = tensor.reshape(batch_size * head_size, seq_len * extra_dim, dim // head_size)
509

Patrick von Platen's avatar
Patrick von Platen committed
510
511
        return tensor

512
513
514
515
516
517
518
519
520
521
522
523
524
525
    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.
        """
Patrick von Platen's avatar
Patrick von Platen committed
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
        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,
        )
547
        del baddbmm_input
Patrick von Platen's avatar
Patrick von Platen committed
548
549
550
551
552

        if self.upcast_softmax:
            attention_scores = attention_scores.float()

        attention_probs = attention_scores.softmax(dim=-1)
553
554
        del attention_scores

Patrick von Platen's avatar
Patrick von Platen committed
555
556
557
558
        attention_probs = attention_probs.to(dtype)

        return attention_probs

559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
    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.
        """
Patrick von Platen's avatar
Patrick von Platen committed
578
579
580
581
        head_size = self.heads
        if attention_mask is None:
            return attention_mask

582
        current_length: int = attention_mask.shape[-1]
583
        if current_length != target_length:
Patrick von Platen's avatar
Patrick von Platen committed
584
585
586
587
588
589
590
            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:
591
592
593
594
                # 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
Patrick von Platen's avatar
Patrick von Platen committed
595
596
                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)

597
598
599
600
601
602
603
        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)

Patrick von Platen's avatar
Patrick von Platen committed
604
605
        return attention_mask

606
607
608
609
610
611
612
613
614
615
616
    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.
        """
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
        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

635
636
637
638
639
    @torch.no_grad()
    def fuse_projections(self, fuse=True):
        device = self.to_q.weight.data.device
        dtype = self.to_q.weight.data.dtype

640
        if not self.is_cross_attention:
641
642
643
644
645
646
            # 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.
647
            self.to_qkv = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
648
            self.to_qkv.weight.copy_(concatenated_weights)
649
650
651
            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)
652
653
654
655
656
657

        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]

658
            self.to_kv = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
659
            self.to_kv.weight.copy_(concatenated_weights)
660
661
662
            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)
663
664
665

        self.fused_projections = fuse

Patrick von Platen's avatar
Patrick von Platen committed
666
667

class AttnProcessor:
668
669
670
671
    r"""
    Default processor for performing attention-related computations.
    """

Patrick von Platen's avatar
Patrick von Platen committed
672
673
674
    def __call__(
        self,
        attn: Attention,
675
676
677
678
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
679
680
        *args,
        **kwargs,
681
    ) -> torch.Tensor:
682
683
684
        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)
685

686
        residual = hidden_states
687

YiYi Xu's avatar
YiYi Xu committed
688
689
690
        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

691
692
693
694
695
696
        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)

Patrick von Platen's avatar
Patrick von Platen committed
697
698
699
700
        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)
701
702
703
704

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

705
        query = attn.to_q(hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
706
707
708

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
709
710
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
711

712
713
        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)
714

Patrick von Platen's avatar
Patrick von Platen committed
715
716
717
718
719
720
721
722
723
        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
724
        hidden_states = attn.to_out[0](hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
725
726
727
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

728
729
730
731
732
733
734
735
        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

Patrick von Platen's avatar
Patrick von Platen committed
736
737
738
        return hidden_states


739
class CustomDiffusionAttnProcessor(nn.Module):
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
    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.
    """

758
759
    def __init__(
        self,
760
761
762
763
764
765
        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,
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
    ):
        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))

784
785
786
    def __call__(
        self,
        attn: Attention,
787
788
789
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
790
    ) -> torch.Tensor:
791
792
793
        batch_size, sequence_length, _ = hidden_states.shape
        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
        if self.train_q_out:
794
            query = self.to_q_custom_diffusion(hidden_states).to(attn.to_q.weight.dtype)
795
        else:
796
            query = attn.to_q(hidden_states.to(attn.to_q.weight.dtype))
797
798
799
800
801
802
803
804
805
806

        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:
807
808
809
810
            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)
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
        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


Patrick von Platen's avatar
Patrick von Platen committed
843
class AttnAddedKVProcessor:
844
845
846
847
848
    r"""
    Processor for performing attention-related computations with extra learnable key and value matrices for the text
    encoder.
    """

849
850
851
    def __call__(
        self,
        attn: Attention,
852
853
854
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
855
856
        *args,
        **kwargs,
857
    ) -> torch.Tensor:
858
859
860
        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)
861

862
        residual = hidden_states
863

Patrick von Platen's avatar
Patrick von Platen committed
864
865
866
867
868
        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)

869
870
871
872
873
        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)

Patrick von Platen's avatar
Patrick von Platen committed
874
875
        hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

876
        query = attn.to_q(hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
877
878
        query = attn.head_to_batch_dim(query)

879
880
        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
881
882
883
        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)

884
        if not attn.only_cross_attention:
885
886
            key = attn.to_k(hidden_states)
            value = attn.to_v(hidden_states)
887
888
889
890
891
892
893
            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
Patrick von Platen's avatar
Patrick von Platen committed
894
895
896
897
898
899

        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
900
        hidden_states = attn.to_out[0](hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
901
902
903
904
905
906
907
908
909
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
        hidden_states = hidden_states + residual

        return hidden_states


910
class AttnAddedKVProcessor2_0:
911
912
913
914
915
    r"""
    Processor for performing scaled dot-product attention (enabled by default if you're using PyTorch 2.0), with extra
    learnable key and value matrices for the text encoder.
    """

916
917
918
919
920
921
    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError(
                "AttnAddedKVProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
            )

922
923
924
    def __call__(
        self,
        attn: Attention,
925
926
927
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
928
929
        *args,
        **kwargs,
930
    ) -> torch.Tensor:
931
932
933
        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)
934

935
        residual = hidden_states
936

937
938
939
940
941
942
943
944
945
946
947
948
        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, out_dim=4)

        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)

949
        query = attn.to_q(hidden_states)
950
951
952
953
954
955
956
957
        query = attn.head_to_batch_dim(query, out_dim=4)

        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, out_dim=4)
        encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj, out_dim=4)

        if not attn.only_cross_attention:
958
959
            key = attn.to_k(hidden_states)
            value = attn.to_v(hidden_states)
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
            key = attn.head_to_batch_dim(key, out_dim=4)
            value = attn.head_to_batch_dim(value, out_dim=4)
            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
        else:
            key = encoder_hidden_states_key_proj
            value = encoder_hidden_states_value_proj

        # 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, residual.shape[1])

        # linear proj
976
        hidden_states = attn.to_out[0](hidden_states)
Will Berman's avatar
Will Berman committed
977
978
979
980
981
982
983
984
985
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
        hidden_states = hidden_states + residual

        return hidden_states


Dhruv Nair's avatar
Dhruv Nair committed
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
class JointAttnProcessor2_0:
    """Attention processor used typically in processing the SD3-like self-attention projections."""

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: torch.FloatTensor = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        *args,
        **kwargs,
    ) -> torch.FloatTensor:
        residual = hidden_states

        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)
        context_input_ndim = encoder_hidden_states.ndim
        if context_input_ndim == 4:
            batch_size, channel, height, width = encoder_hidden_states.shape
            encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size = encoder_hidden_states.shape[0]

        # `sample` projections.
        query = attn.to_q(hidden_states)
        key = attn.to_k(hidden_states)
        value = attn.to_v(hidden_states)

        # `context` projections.
        encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)

        # attention
        query = torch.cat([query, encoder_hidden_states_query_proj], dim=1)
        key = torch.cat([key, encoder_hidden_states_key_proj], dim=1)
        value = torch.cat([value, encoder_hidden_states_value_proj], dim=1)

        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)

1036
        hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
Dhruv Nair's avatar
Dhruv Nair committed
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states = hidden_states.to(query.dtype)

        # Split the attention outputs.
        hidden_states, encoder_hidden_states = (
            hidden_states[:, : residual.shape[1]],
            hidden_states[:, residual.shape[1] :],
        )

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)
        if not attn.context_pre_only:
            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
        if context_input_ndim == 4:
            encoder_hidden_states = encoder_hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        return hidden_states, encoder_hidden_states


class FusedJointAttnProcessor2_0:
    """Attention processor used typically in processing the SD3-like self-attention projections."""

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: torch.FloatTensor = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        *args,
        **kwargs,
    ) -> torch.FloatTensor:
        residual = hidden_states

        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)
        context_input_ndim = encoder_hidden_states.ndim
        if context_input_ndim == 4:
            batch_size, channel, height, width = encoder_hidden_states.shape
            encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size = encoder_hidden_states.shape[0]

        # `sample` projections.
        qkv = attn.to_qkv(hidden_states)
        split_size = qkv.shape[-1] // 3
        query, key, value = torch.split(qkv, split_size, dim=-1)

        # `context` projections.
        encoder_qkv = attn.to_added_qkv(encoder_hidden_states)
        split_size = encoder_qkv.shape[-1] // 3
        (
            encoder_hidden_states_query_proj,
            encoder_hidden_states_key_proj,
            encoder_hidden_states_value_proj,
        ) = torch.split(encoder_qkv, split_size, dim=-1)

        # attention
        query = torch.cat([query, encoder_hidden_states_query_proj], dim=1)
        key = torch.cat([key, encoder_hidden_states_key_proj], dim=1)
        value = torch.cat([value, encoder_hidden_states_value_proj], dim=1)

        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)

1115
        hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
Dhruv Nair's avatar
Dhruv Nair committed
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states = hidden_states.to(query.dtype)

        # Split the attention outputs.
        hidden_states, encoder_hidden_states = (
            hidden_states[:, : residual.shape[1]],
            hidden_states[:, residual.shape[1] :],
        )

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)
        if not attn.context_pre_only:
            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
        if context_input_ndim == 4:
            encoder_hidden_states = encoder_hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        return hidden_states, encoder_hidden_states


1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
class XFormersAttnAddedKVProcessor:
    r"""
    Processor for implementing memory efficient attention using xFormers.

    Args:
        attention_op (`Callable`, *optional*, defaults to `None`):
            The base
            [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
            use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
            operator.
    """

    def __init__(self, attention_op: Optional[Callable] = None):
        self.attention_op = attention_op

1155
1156
1157
    def __call__(
        self,
        attn: Attention,
1158
1159
1160
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
1161
    ) -> torch.Tensor:
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
        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

        hidden_states = xformers.ops.memory_efficient_attention(
            query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
        )
        hidden_states = hidden_states.to(query.dtype)
        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)

        hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
        hidden_states = hidden_states + residual

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
1211
class XFormersAttnProcessor:
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
    r"""
    Processor for implementing memory efficient attention using xFormers.

    Args:
        attention_op (`Callable`, *optional*, defaults to `None`):
            The base
            [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
            use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
            operator.
    """

Patrick von Platen's avatar
Patrick von Platen committed
1223
1224
1225
    def __init__(self, attention_op: Optional[Callable] = None):
        self.attention_op = attention_op

1226
1227
1228
    def __call__(
        self,
        attn: Attention,
1229
1230
1231
1232
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
1233
1234
        *args,
        **kwargs,
1235
    ) -> torch.Tensor:
1236
1237
1238
        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)
1239

1240
        residual = hidden_states
1241

1242
1243
1244
        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

1245
1246
1247
1248
1249
1250
        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)

1251
        batch_size, key_tokens, _ = (
Patrick von Platen's avatar
Patrick von Platen committed
1252
1253
1254
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )

1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
        attention_mask = attn.prepare_attention_mask(attention_mask, key_tokens, batch_size)
        if attention_mask is not None:
            # expand our mask's singleton query_tokens dimension:
            #   [batch*heads,            1, key_tokens] ->
            #   [batch*heads, query_tokens, key_tokens]
            # so that it can be added as a bias onto the attention scores that xformers computes:
            #   [batch*heads, query_tokens, key_tokens]
            # we do this explicitly because xformers doesn't broadcast the singleton dimension for us.
            _, query_tokens, _ = hidden_states.shape
            attention_mask = attention_mask.expand(-1, query_tokens, -1)
Patrick von Platen's avatar
Patrick von Platen committed
1265

1266
1267
1268
        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

1269
        query = attn.to_q(hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
1270
1271
1272

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
1273
1274
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
1275

1276
1277
        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289

        query = attn.head_to_batch_dim(query).contiguous()
        key = attn.head_to_batch_dim(key).contiguous()
        value = attn.head_to_batch_dim(value).contiguous()

        hidden_states = xformers.ops.memory_efficient_attention(
            query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
        )
        hidden_states = hidden_states.to(query.dtype)
        hidden_states = attn.batch_to_head_dim(hidden_states)

        # linear proj
1290
        hidden_states = attn.to_out[0](hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
1291
1292
        # dropout
        hidden_states = attn.to_out[1](hidden_states)
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319

        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 AttnProcessorNPU:
    r"""
    Processor for implementing flash attention using torch_npu. Torch_npu supports only fp16 and bf16 data types. If
    fp32 is used, F.scaled_dot_product_attention will be used for computation, but the acceleration effect on NPU is
    not significant.

    """

    def __init__(self):
        if not is_torch_npu_available():
            raise ImportError("AttnProcessorNPU requires torch_npu extensions and is supported only on npu devices.")

    def __call__(
        self,
        attn: Attention,
1320
1321
1322
1323
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
1324
1325
        *args,
        **kwargs,
1326
    ) -> torch.Tensor:
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
        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
        )

        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)

        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)
        if query.dtype in (torch.float16, torch.bfloat16):
            hidden_states = torch_npu.npu_fusion_attention(
                query,
                key,
                value,
                attn.heads,
                input_layout="BNSD",
                pse=None,
                atten_mask=attention_mask,
                scale=1.0 / math.sqrt(query.shape[-1]),
                pre_tockens=65536,
                next_tockens=65536,
                keep_prob=1.0,
                sync=False,
                inner_precise=0,
            )[0]
        else:
            # 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)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)
1402
1403
1404
1405
1406
1407
1408
1409
1410

        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

Patrick von Platen's avatar
Patrick von Platen committed
1411
1412
1413
1414
        return hidden_states


class AttnProcessor2_0:
1415
1416
1417
1418
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
    """

Patrick von Platen's avatar
Patrick von Platen committed
1419
1420
1421
1422
    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

YiYi Xu's avatar
YiYi Xu committed
1423
1424
1425
    def __call__(
        self,
        attn: Attention,
1426
1427
1428
1429
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
1430
1431
        *args,
        **kwargs,
1432
    ) -> torch.Tensor:
1433
1434
1435
1436
        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)

1437
        residual = hidden_states
YiYi Xu's avatar
YiYi Xu committed
1438
1439
1440
        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

1441
1442
1443
1444
1445
1446
        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)

Patrick von Platen's avatar
Patrick von Platen committed
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
        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])

1457
1458
1459
        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

1460
        query = attn.to_q(hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
1461
1462
1463

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
1464
1465
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
1466

1467
1468
        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
1469

1470
        inner_dim = key.shape[-1]
Patrick von Platen's avatar
Patrick von Platen committed
1471
        head_dim = inner_dim // attn.heads
1472

Patrick von Platen's avatar
Patrick von Platen committed
1473
        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
1474

Patrick von Platen's avatar
Patrick von Platen committed
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
        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)

        # linear proj
1488
        hidden_states = attn.to_out[0](hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
1489
1490
        # dropout
        hidden_states = attn.to_out[1](hidden_states)
1491
1492
1493
1494
1495

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if attn.residual_connection:
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class HunyuanAttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
    used in the HunyuanDiT model. It applies a s normalization layer and rotary embedding on query and key vector.
    """

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

    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,
        image_rotary_emb: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        from .embeddings import apply_rotary_emb

        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)

        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)

        if attn.norm_q is not None:
            query = attn.norm_q(query)
        if attn.norm_k is not None:
            key = attn.norm_k(key)

        # Apply RoPE if needed
        if image_rotary_emb is not None:
            query = apply_rotary_emb(query, image_rotary_emb)
            if not attn.is_cross_attention:
                key = apply_rotary_emb(key, image_rotary_emb)

        # 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)

        # 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:
1594
1595
1596
1597
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

Patrick von Platen's avatar
Patrick von Platen committed
1598
1599
1600
        return hidden_states


1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
class LuminaAttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
    used in the LuminaNextDiT model. It applies a s normalization layer and rotary embedding on query and key vector.
    """

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        query_rotary_emb: Optional[torch.Tensor] = None,
        key_rotary_emb: Optional[torch.Tensor] = None,
        base_sequence_length: Optional[int] = None,
    ) -> torch.Tensor:
        from .embeddings import apply_rotary_emb

        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

        # Get Query-Key-Value Pair
        query = attn.to_q(hidden_states)
        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        query_dim = query.shape[-1]
        inner_dim = key.shape[-1]
        head_dim = query_dim // attn.heads
        dtype = query.dtype

        # Get key-value heads
        kv_heads = inner_dim // head_dim

        # Apply Query-Key Norm if needed
        if attn.norm_q is not None:
            query = attn.norm_q(query)
        if attn.norm_k is not None:
            key = attn.norm_k(key)

        query = query.view(batch_size, -1, attn.heads, head_dim)

        key = key.view(batch_size, -1, kv_heads, head_dim)
        value = value.view(batch_size, -1, kv_heads, head_dim)

        # Apply RoPE if needed
        if query_rotary_emb is not None:
            query = apply_rotary_emb(query, query_rotary_emb, use_real=False)
        if key_rotary_emb is not None:
            key = apply_rotary_emb(key, key_rotary_emb, use_real=False)

        query, key = query.to(dtype), key.to(dtype)

        # Apply proportional attention if true
        if key_rotary_emb is None:
            softmax_scale = None
        else:
            if base_sequence_length is not None:
                softmax_scale = math.sqrt(math.log(sequence_length, base_sequence_length)) * attn.scale
            else:
                softmax_scale = attn.scale

        # perform Grouped-qurey Attention (GQA)
        n_rep = attn.heads // kv_heads
        if n_rep >= 1:
            key = key.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
            value = value.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)

        # scaled_dot_product_attention expects attention_mask shape to be
        # (batch, heads, source_length, target_length)
        attention_mask = attention_mask.bool().view(batch_size, 1, 1, -1)
        attention_mask = attention_mask.expand(-1, attn.heads, sequence_length, -1)

        query = query.transpose(1, 2)
        key = key.transpose(1, 2)
        value = value.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, scale=softmax_scale
        )
        hidden_states = hidden_states.transpose(1, 2).to(dtype)

        return hidden_states


1697
1698
class FusedAttnProcessor2_0:
    r"""
1699
1700
1701
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). It uses
    fused projection layers. 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.
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718

    <Tip warning={true}>

    This API is currently 🧪 experimental in nature and can change in future.

    </Tip>
    """

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError(
                "FusedAttnProcessor2_0 requires at least PyTorch 2.0, to use it. Please upgrade PyTorch to > 2.0."
            )

    def __call__(
        self,
        attn: Attention,
1719
1720
1721
1722
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
1723
1724
        *args,
        **kwargs,
1725
    ) -> torch.Tensor:
1726
1727
1728
1729
        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)

1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
        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)

        if encoder_hidden_states is None:
1754
            qkv = attn.to_qkv(hidden_states)
1755
1756
1757
1758
1759
            split_size = qkv.shape[-1] // 3
            query, key, value = torch.split(qkv, split_size, dim=-1)
        else:
            if attn.norm_cross:
                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
1760
            query = attn.to_q(hidden_states)
1761

1762
            kv = attn.to_kv(encoder_hidden_states)
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
            split_size = kv.shape[-1] // 2
            key, value = torch.split(kv, split_size, dim=-1)

        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)

        # linear proj
1783
        hidden_states = attn.to_out[0](hidden_states)
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
        # 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


1798
class CustomDiffusionXFormersAttnProcessor(nn.Module):
1799
    r"""
1800
    Processor for implementing memory efficient attention using xFormers for the Custom Diffusion method.
1801
1802

    Args:
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
    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.
    attention_op (`Callable`, *optional*, defaults to `None`):
        The base
        [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to use
        as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best operator.
1819
1820
    """

1821
    def __init__(
1822
        self,
1823
1824
1825
1826
1827
1828
        train_kv: bool = True,
        train_q_out: bool = False,
        hidden_size: Optional[int] = None,
        cross_attention_dim: Optional[int] = None,
        out_bias: bool = True,
        dropout: float = 0.0,
1829
        attention_op: Optional[Callable] = None,
1830
    ):
Patrick von Platen's avatar
Patrick von Platen committed
1831
        super().__init__()
1832
1833
        self.train_kv = train_kv
        self.train_q_out = train_q_out
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848

        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.attention_op = attention_op

        # `_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))

1849
1850
1851
    def __call__(
        self,
        attn: Attention,
1852
1853
1854
1855
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
1856
1857
1858
1859
1860
1861
1862
        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 self.train_q_out:
1863
            query = self.to_q_custom_diffusion(hidden_states).to(attn.to_q.weight.dtype)
1864
        else:
1865
            query = attn.to_q(hidden_states.to(attn.to_q.weight.dtype))
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875

        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:
1876
1877
1878
1879
            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)
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
        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).contiguous()
        key = attn.head_to_batch_dim(key).contiguous()
        value = attn.head_to_batch_dim(value).contiguous()

        hidden_states = xformers.ops.memory_efficient_attention(
            query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
        )
        hidden_states = hidden_states.to(query.dtype)
        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)
1910

1911
1912
1913
        return hidden_states


1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
class CustomDiffusionAttnProcessor2_0(nn.Module):
    r"""
    Processor for implementing attention for the Custom Diffusion method using PyTorch 2.0’s memory-efficient scaled
    dot-product attention.

    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,
1936
1937
1938
1939
1940
1941
        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,
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
    ):
        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))

1960
1961
1962
    def __call__(
        self,
        attn: Attention,
1963
1964
1965
1966
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
        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)
        else:
            query = attn.to_q(hidden_states)

        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:
1983
1984
1985
1986
1987
            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)

1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
        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()

        inner_dim = hidden_states.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)

        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


Patrick von Platen's avatar
Patrick von Platen committed
2028
class SlicedAttnProcessor:
2029
2030
2031
2032
2033
2034
2035
2036
2037
    r"""
    Processor for implementing sliced attention.

    Args:
        slice_size (`int`, *optional*):
            The number of steps to compute attention. Uses as many slices as `attention_head_dim // slice_size`, and
            `attention_head_dim` must be a multiple of the `slice_size`.
    """

2038
    def __init__(self, slice_size: int):
Patrick von Platen's avatar
Patrick von Platen committed
2039
2040
        self.slice_size = slice_size

2041
2042
2043
    def __call__(
        self,
        attn: Attention,
2044
2045
2046
2047
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
2048
2049
2050
2051
2052
2053
2054
2055
        residual = hidden_states

        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)

Patrick von Platen's avatar
Patrick von Platen committed
2056
2057
2058
2059
2060
        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)

2061
2062
2063
        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

Patrick von Platen's avatar
Patrick von Platen committed
2064
2065
2066
2067
2068
2069
        query = attn.to_q(hidden_states)
        dim = query.shape[-1]
        query = attn.head_to_batch_dim(query)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
2070
2071
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)
        key = attn.head_to_batch_dim(key)
        value = attn.head_to_batch_dim(value)

        batch_size_attention, query_tokens, _ = query.shape
        hidden_states = torch.zeros(
            (batch_size_attention, query_tokens, dim // attn.heads), device=query.device, dtype=query.dtype
        )

        for i in range(batch_size_attention // self.slice_size):
            start_idx = i * self.slice_size
            end_idx = (i + 1) * self.slice_size

            query_slice = query[start_idx:end_idx]
            key_slice = key[start_idx:end_idx]
            attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None

            attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)

            attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])

            hidden_states[start_idx:end_idx] = attn_slice

        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)

2104
2105
2106
2107
2108
2109
2110
2111
        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

Patrick von Platen's avatar
Patrick von Platen committed
2112
2113
2114
2115
        return hidden_states


class SlicedAttnAddedKVProcessor:
2116
2117
2118
2119
2120
2121
2122
2123
2124
    r"""
    Processor for implementing sliced attention with extra learnable key and value matrices for the text encoder.

    Args:
        slice_size (`int`, *optional*):
            The number of steps to compute attention. Uses as many slices as `attention_head_dim // slice_size`, and
            `attention_head_dim` must be a multiple of the `slice_size`.
    """

Patrick von Platen's avatar
Patrick von Platen committed
2125
2126
2127
    def __init__(self, slice_size):
        self.slice_size = slice_size

2128
2129
2130
    def __call__(
        self,
        attn: "Attention",
2131
2132
2133
2134
2135
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
Patrick von Platen's avatar
Patrick von Platen committed
2136
        residual = hidden_states
2137
2138
2139
2140

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

Patrick von Platen's avatar
Patrick von Platen committed
2141
2142
2143
2144
2145
2146
        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)

2147
2148
2149
2150
2151
        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)

Patrick von Platen's avatar
Patrick von Platen committed
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
        hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states)
        dim = query.shape[-1]
        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)

2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
        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
Patrick von Platen's avatar
Patrick von Platen committed
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206

        batch_size_attention, query_tokens, _ = query.shape
        hidden_states = torch.zeros(
            (batch_size_attention, query_tokens, dim // attn.heads), device=query.device, dtype=query.dtype
        )

        for i in range(batch_size_attention // self.slice_size):
            start_idx = i * self.slice_size
            end_idx = (i + 1) * self.slice_size

            query_slice = query[start_idx:end_idx]
            key_slice = key[start_idx:end_idx]
            attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None

            attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)

            attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])

            hidden_states[start_idx:end_idx] = attn_slice

        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)

        hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
        hidden_states = hidden_states + residual

        return hidden_states


YiYi Xu's avatar
YiYi Xu committed
2207
2208
class SpatialNorm(nn.Module):
    """
2209
2210
2211
2212
2213
2214
2215
    Spatially conditioned normalization as defined in https://arxiv.org/abs/2209.09002.

    Args:
        f_channels (`int`):
            The number of channels for input to group normalization layer, and output of the spatial norm layer.
        zq_channels (`int`):
            The number of channels for the quantized vector as described in the paper.
YiYi Xu's avatar
YiYi Xu committed
2216
2217
2218
2219
    """

    def __init__(
        self,
2220
2221
        f_channels: int,
        zq_channels: int,
YiYi Xu's avatar
YiYi Xu committed
2222
2223
2224
2225
2226
2227
    ):
        super().__init__()
        self.norm_layer = nn.GroupNorm(num_channels=f_channels, num_groups=32, eps=1e-6, affine=True)
        self.conv_y = nn.Conv2d(zq_channels, f_channels, kernel_size=1, stride=1, padding=0)
        self.conv_b = nn.Conv2d(zq_channels, f_channels, kernel_size=1, stride=1, padding=0)

2228
    def forward(self, f: torch.Tensor, zq: torch.Tensor) -> torch.Tensor:
YiYi Xu's avatar
YiYi Xu committed
2229
2230
2231
2232
2233
        f_size = f.shape[-2:]
        zq = F.interpolate(zq, size=f_size, mode="nearest")
        norm_f = self.norm_layer(f)
        new_f = norm_f * self.conv_y(zq) + self.conv_b(zq)
        return new_f
2234
2235


2236
2237
class IPAdapterAttnProcessor(nn.Module):
    r"""
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
2238
    Attention processor for Multiple IP-Adapters.
2239
2240
2241
2242
2243
2244

    Args:
        hidden_size (`int`):
            The hidden size of the attention layer.
        cross_attention_dim (`int`):
            The number of channels in the `encoder_hidden_states`.
2245
        num_tokens (`int`, `Tuple[int]` or `List[int]`, defaults to `(4,)`):
2246
            The context length of the image features.
2247
        scale (`float` or List[`float`], defaults to 1.0):
2248
2249
2250
            the weight scale of image prompt.
    """

2251
    def __init__(self, hidden_size, cross_attention_dim=None, num_tokens=(4,), scale=1.0):
2252
2253
2254
2255
        super().__init__()

        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
2256
2257
2258

        if not isinstance(num_tokens, (tuple, list)):
            num_tokens = [num_tokens]
2259
        self.num_tokens = num_tokens
2260
2261
2262
2263
2264

        if not isinstance(scale, list):
            scale = [scale] * len(num_tokens)
        if len(scale) != len(num_tokens):
            raise ValueError("`scale` should be a list of integers with the same length as `num_tokens`.")
2265
2266
        self.scale = scale

2267
2268
2269
2270
2271
2272
        self.to_k_ip = nn.ModuleList(
            [nn.Linear(cross_attention_dim, hidden_size, bias=False) for _ in range(len(num_tokens))]
        )
        self.to_v_ip = nn.ModuleList(
            [nn.Linear(cross_attention_dim, hidden_size, bias=False) for _ in range(len(num_tokens))]
        )
2273
2274
2275

    def __call__(
        self,
2276
        attn: Attention,
2277
2278
2279
2280
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
2281
        scale: float = 1.0,
2282
        ip_adapter_masks: Optional[torch.Tensor] = None,
2283
2284
2285
    ):
        residual = hidden_states

2286
2287
2288
2289
2290
2291
        # separate ip_hidden_states from encoder_hidden_states
        if encoder_hidden_states is not None:
            if isinstance(encoder_hidden_states, tuple):
                encoder_hidden_states, ip_hidden_states = encoder_hidden_states
            else:
                deprecation_message = (
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
2292
2293
                    "You have passed a tensor as `encoder_hidden_states`. This is deprecated and will be removed in a future release."
                    " Please make sure to update your script to pass `encoder_hidden_states` as a tuple to suppress this warning."
2294
2295
2296
2297
2298
2299
2300
2301
                )
                deprecate("encoder_hidden_states not a tuple", "1.0.0", deprecation_message, standard_warn=False)
                end_pos = encoder_hidden_states.shape[1] - self.num_tokens[0]
                encoder_hidden_states, ip_hidden_states = (
                    encoder_hidden_states[:, :end_pos, :],
                    [encoder_hidden_states[:, end_pos:, :]],
                )

2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
        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)

2337
        if ip_adapter_masks is not None:
2338
2339
2340
2341
            if not isinstance(ip_adapter_masks, List):
                # for backward compatibility, we accept `ip_adapter_mask` as a tensor of shape [num_ip_adapter, 1, height, width]
                ip_adapter_masks = list(ip_adapter_masks.unsqueeze(1))
            if not (len(ip_adapter_masks) == len(self.scale) == len(ip_hidden_states)):
2342
                raise ValueError(
2343
2344
2345
                    f"Length of ip_adapter_masks array ({len(ip_adapter_masks)}) must match "
                    f"length of self.scale array ({len(self.scale)}) and number of ip_hidden_states "
                    f"({len(ip_hidden_states)})"
2346
                )
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
            else:
                for index, (mask, scale, ip_state) in enumerate(zip(ip_adapter_masks, self.scale, ip_hidden_states)):
                    if not isinstance(mask, torch.Tensor) or mask.ndim != 4:
                        raise ValueError(
                            "Each element of the ip_adapter_masks array should be a tensor with shape "
                            "[1, num_images_for_ip_adapter, height, width]."
                            " Please use `IPAdapterMaskProcessor` to preprocess your mask"
                        )
                    if mask.shape[1] != ip_state.shape[1]:
                        raise ValueError(
                            f"Number of masks ({mask.shape[1]}) does not match "
                            f"number of ip images ({ip_state.shape[1]}) at index {index}"
                        )
                    if isinstance(scale, list) and not len(scale) == mask.shape[1]:
                        raise ValueError(
                            f"Number of masks ({mask.shape[1]}) does not match "
                            f"number of scales ({len(scale)}) at index {index}"
                        )
2365
2366
2367
        else:
            ip_adapter_masks = [None] * len(self.scale)

2368
        # for ip-adapter
2369
2370
        for current_ip_hidden_states, scale, to_k_ip, to_v_ip, mask in zip(
            ip_hidden_states, self.scale, self.to_k_ip, self.to_v_ip, ip_adapter_masks
2371
        ):
Jenyuan-Huang's avatar
Jenyuan-Huang committed
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
            skip = False
            if isinstance(scale, list):
                if all(s == 0 for s in scale):
                    skip = True
            elif scale == 0:
                skip = True
            if not skip:
                if mask is not None:
                    if not isinstance(scale, list):
                        scale = [scale] * mask.shape[1]

                    current_num_images = mask.shape[1]
                    for i in range(current_num_images):
                        ip_key = to_k_ip(current_ip_hidden_states[:, i, :, :])
                        ip_value = to_v_ip(current_ip_hidden_states[:, i, :, :])

                        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)
                        _current_ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
                        _current_ip_hidden_states = attn.batch_to_head_dim(_current_ip_hidden_states)

                        mask_downsample = IPAdapterMaskProcessor.downsample(
                            mask[:, i, :, :],
                            batch_size,
                            _current_ip_hidden_states.shape[1],
                            _current_ip_hidden_states.shape[2],
                        )

                        mask_downsample = mask_downsample.to(dtype=query.dtype, device=query.device)
2403

Jenyuan-Huang's avatar
Jenyuan-Huang committed
2404
2405
2406
2407
                        hidden_states = hidden_states + scale[i] * (_current_ip_hidden_states * mask_downsample)
                else:
                    ip_key = to_k_ip(current_ip_hidden_states)
                    ip_value = to_v_ip(current_ip_hidden_states)
2408
2409
2410
2411
2412

                    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)
Jenyuan-Huang's avatar
Jenyuan-Huang committed
2413
2414
                    current_ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
                    current_ip_hidden_states = attn.batch_to_head_dim(current_ip_hidden_states)
2415

Jenyuan-Huang's avatar
Jenyuan-Huang committed
2416
                    hidden_states = hidden_states + scale * current_ip_hidden_states
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435

        # 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"""
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
2436
    Attention processor for IP-Adapter for PyTorch 2.0.
2437
2438
2439
2440
2441
2442

    Args:
        hidden_size (`int`):
            The hidden size of the attention layer.
        cross_attention_dim (`int`):
            The number of channels in the `encoder_hidden_states`.
2443
        num_tokens (`int`, `Tuple[int]` or `List[int]`, defaults to `(4,)`):
2444
            The context length of the image features.
2445
        scale (`float` or `List[float]`, defaults to 1.0):
2446
2447
2448
            the weight scale of image prompt.
    """

2449
    def __init__(self, hidden_size, cross_attention_dim=None, num_tokens=(4,), scale=1.0):
2450
2451
2452
2453
2454
2455
2456
2457
2458
        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
2459
2460
2461

        if not isinstance(num_tokens, (tuple, list)):
            num_tokens = [num_tokens]
2462
        self.num_tokens = num_tokens
2463
2464
2465
2466
2467

        if not isinstance(scale, list):
            scale = [scale] * len(num_tokens)
        if len(scale) != len(num_tokens):
            raise ValueError("`scale` should be a list of integers with the same length as `num_tokens`.")
2468
2469
        self.scale = scale

2470
2471
2472
2473
2474
2475
        self.to_k_ip = nn.ModuleList(
            [nn.Linear(cross_attention_dim, hidden_size, bias=False) for _ in range(len(num_tokens))]
        )
        self.to_v_ip = nn.ModuleList(
            [nn.Linear(cross_attention_dim, hidden_size, bias=False) for _ in range(len(num_tokens))]
        )
2476
2477
2478

    def __call__(
        self,
2479
        attn: Attention,
2480
2481
2482
2483
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
2484
        scale: float = 1.0,
2485
        ip_adapter_masks: Optional[torch.Tensor] = None,
2486
2487
2488
    ):
        residual = hidden_states

2489
2490
2491
2492
2493
2494
        # separate ip_hidden_states from encoder_hidden_states
        if encoder_hidden_states is not None:
            if isinstance(encoder_hidden_states, tuple):
                encoder_hidden_states, ip_hidden_states = encoder_hidden_states
            else:
                deprecation_message = (
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
2495
2496
                    "You have passed a tensor as `encoder_hidden_states`. This is deprecated and will be removed in a future release."
                    " Please make sure to update your script to pass `encoder_hidden_states` as a tuple to suppress this warning."
2497
2498
2499
2500
2501
2502
2503
2504
                )
                deprecate("encoder_hidden_states not a tuple", "1.0.0", deprecation_message, standard_warn=False)
                end_pos = encoder_hidden_states.shape[1] - self.num_tokens[0]
                encoder_hidden_states, ip_hidden_states = (
                    encoder_hidden_states[:, :end_pos, :],
                    [encoder_hidden_states[:, end_pos:, :]],
                )

2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
        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)

        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)

2554
        if ip_adapter_masks is not None:
2555
2556
2557
2558
            if not isinstance(ip_adapter_masks, List):
                # for backward compatibility, we accept `ip_adapter_mask` as a tensor of shape [num_ip_adapter, 1, height, width]
                ip_adapter_masks = list(ip_adapter_masks.unsqueeze(1))
            if not (len(ip_adapter_masks) == len(self.scale) == len(ip_hidden_states)):
2559
                raise ValueError(
2560
2561
2562
                    f"Length of ip_adapter_masks array ({len(ip_adapter_masks)}) must match "
                    f"length of self.scale array ({len(self.scale)}) and number of ip_hidden_states "
                    f"({len(ip_hidden_states)})"
2563
                )
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
            else:
                for index, (mask, scale, ip_state) in enumerate(zip(ip_adapter_masks, self.scale, ip_hidden_states)):
                    if not isinstance(mask, torch.Tensor) or mask.ndim != 4:
                        raise ValueError(
                            "Each element of the ip_adapter_masks array should be a tensor with shape "
                            "[1, num_images_for_ip_adapter, height, width]."
                            " Please use `IPAdapterMaskProcessor` to preprocess your mask"
                        )
                    if mask.shape[1] != ip_state.shape[1]:
                        raise ValueError(
                            f"Number of masks ({mask.shape[1]}) does not match "
                            f"number of ip images ({ip_state.shape[1]}) at index {index}"
                        )
                    if isinstance(scale, list) and not len(scale) == mask.shape[1]:
                        raise ValueError(
                            f"Number of masks ({mask.shape[1]}) does not match "
                            f"number of scales ({len(scale)}) at index {index}"
                        )
2582
2583
2584
        else:
            ip_adapter_masks = [None] * len(self.scale)

2585
        # for ip-adapter
2586
2587
        for current_ip_hidden_states, scale, to_k_ip, to_v_ip, mask in zip(
            ip_hidden_states, self.scale, self.to_k_ip, self.to_v_ip, ip_adapter_masks
2588
        ):
Jenyuan-Huang's avatar
Jenyuan-Huang committed
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
            skip = False
            if isinstance(scale, list):
                if all(s == 0 for s in scale):
                    skip = True
            elif scale == 0:
                skip = True
            if not skip:
                if mask is not None:
                    if not isinstance(scale, list):
                        scale = [scale] * mask.shape[1]

                    current_num_images = mask.shape[1]
                    for i in range(current_num_images):
                        ip_key = to_k_ip(current_ip_hidden_states[:, i, :, :])
                        ip_value = to_v_ip(current_ip_hidden_states[:, i, :, :])

                        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
                        _current_ip_hidden_states = F.scaled_dot_product_attention(
                            query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
                        )

                        _current_ip_hidden_states = _current_ip_hidden_states.transpose(1, 2).reshape(
                            batch_size, -1, attn.heads * head_dim
                        )
                        _current_ip_hidden_states = _current_ip_hidden_states.to(query.dtype)
2618

Jenyuan-Huang's avatar
Jenyuan-Huang committed
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
                        mask_downsample = IPAdapterMaskProcessor.downsample(
                            mask[:, i, :, :],
                            batch_size,
                            _current_ip_hidden_states.shape[1],
                            _current_ip_hidden_states.shape[2],
                        )

                        mask_downsample = mask_downsample.to(dtype=query.dtype, device=query.device)
                        hidden_states = hidden_states + scale[i] * (_current_ip_hidden_states * mask_downsample)
                else:
                    ip_key = to_k_ip(current_ip_hidden_states)
                    ip_value = to_v_ip(current_ip_hidden_states)
2631

2632
2633
                    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)
2634

2635
2636
                    # the output of sdp = (batch, num_heads, seq_len, head_dim)
                    # TODO: add support for attn.scale when we move to Torch 2.1
Jenyuan-Huang's avatar
Jenyuan-Huang committed
2637
                    current_ip_hidden_states = F.scaled_dot_product_attention(
2638
2639
                        query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
                    )
2640

Jenyuan-Huang's avatar
Jenyuan-Huang committed
2641
                    current_ip_hidden_states = current_ip_hidden_states.transpose(1, 2).reshape(
2642
2643
                        batch_size, -1, attn.heads * head_dim
                    )
Jenyuan-Huang's avatar
Jenyuan-Huang committed
2644
                    current_ip_hidden_states = current_ip_hidden_states.to(query.dtype)
2645

Jenyuan-Huang's avatar
Jenyuan-Huang committed
2646
                    hidden_states = hidden_states + scale * current_ip_hidden_states
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663

        # 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


YiYi Xu's avatar
YiYi Xu committed
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
class PAGIdentitySelfAttnProcessor2_0:
    r"""
    Processor for implementing PAG using scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
    PAG reference: https://arxiv.org/abs/2403.17377
    """

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError(
                "PAGIdentitySelfAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
            )

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        temb: Optional[torch.FloatTensor] = None,
    ) -> torch.Tensor:
        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)

        # chunk
        hidden_states_org, hidden_states_ptb = hidden_states.chunk(2)

        # original path
        batch_size, sequence_length, _ = hidden_states_org.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_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states_org)
        key = attn.to_k(hidden_states_org)
        value = attn.to_v(hidden_states_org)

        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_org = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
        )
        hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states_org = hidden_states_org.to(query.dtype)

        # linear proj
        hidden_states_org = attn.to_out[0](hidden_states_org)
        # dropout
        hidden_states_org = attn.to_out[1](hidden_states_org)

        if input_ndim == 4:
            hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width)

        # perturbed path (identity attention)
        batch_size, sequence_length, _ = hidden_states_ptb.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_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2)

        hidden_states_ptb = attn.to_v(hidden_states_ptb)
        hidden_states_ptb = hidden_states_ptb.to(query.dtype)

        # linear proj
        hidden_states_ptb = attn.to_out[0](hidden_states_ptb)
        # dropout
        hidden_states_ptb = attn.to_out[1](hidden_states_ptb)

        if input_ndim == 4:
            hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width)

        # cat
        hidden_states = torch.cat([hidden_states_org, hidden_states_ptb])

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class PAGCFGIdentitySelfAttnProcessor2_0:
    r"""
    Processor for implementing PAG using scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
    PAG reference: https://arxiv.org/abs/2403.17377
    """

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError(
                "PAGCFGIdentitySelfAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
            )

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        temb: Optional[torch.FloatTensor] = None,
    ) -> torch.Tensor:
        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)

        # chunk
        hidden_states_uncond, hidden_states_org, hidden_states_ptb = hidden_states.chunk(3)
        hidden_states_org = torch.cat([hidden_states_uncond, hidden_states_org])

        # original path
        batch_size, sequence_length, _ = hidden_states_org.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_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states_org)
        key = attn.to_k(hidden_states_org)
        value = attn.to_v(hidden_states_org)

        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_org = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
        )

        hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states_org = hidden_states_org.to(query.dtype)

        # linear proj
        hidden_states_org = attn.to_out[0](hidden_states_org)
        # dropout
        hidden_states_org = attn.to_out[1](hidden_states_org)

        if input_ndim == 4:
            hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width)

        # perturbed path (identity attention)
        batch_size, sequence_length, _ = hidden_states_ptb.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_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2)

        value = attn.to_v(hidden_states_ptb)
        hidden_states_ptb = value
        hidden_states_ptb = hidden_states_ptb.to(query.dtype)

        # linear proj
        hidden_states_ptb = attn.to_out[0](hidden_states_ptb)
        # dropout
        hidden_states_ptb = attn.to_out[1](hidden_states_ptb)

        if input_ndim == 4:
            hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width)

        # cat
        hidden_states = torch.cat([hidden_states_org, hidden_states_ptb])

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
class LoRAAttnProcessor:
    def __init__(self):
        pass


class LoRAAttnProcessor2_0:
    def __init__(self):
        pass


class LoRAXFormersAttnProcessor:
    def __init__(self):
        pass


class LoRAAttnAddedKVProcessor:
    def __init__(self):
        pass


2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
ADDED_KV_ATTENTION_PROCESSORS = (
    AttnAddedKVProcessor,
    SlicedAttnAddedKVProcessor,
    AttnAddedKVProcessor2_0,
    XFormersAttnAddedKVProcessor,
)

CROSS_ATTENTION_PROCESSORS = (
    AttnProcessor,
    AttnProcessor2_0,
    XFormersAttnProcessor,
    SlicedAttnProcessor,
2910
2911
    IPAdapterAttnProcessor,
    IPAdapterAttnProcessor2_0,
2912
2913
)

2914
2915
2916
AttentionProcessor = Union[
    AttnProcessor,
    AttnProcessor2_0,
2917
    FusedAttnProcessor2_0,
2918
2919
2920
2921
2922
2923
2924
2925
    XFormersAttnProcessor,
    SlicedAttnProcessor,
    AttnAddedKVProcessor,
    SlicedAttnAddedKVProcessor,
    AttnAddedKVProcessor2_0,
    XFormersAttnAddedKVProcessor,
    CustomDiffusionAttnProcessor,
    CustomDiffusionXFormersAttnProcessor,
2926
    CustomDiffusionAttnProcessor2_0,
YiYi Xu's avatar
YiYi Xu committed
2927
2928
    PAGCFGIdentitySelfAttnProcessor2_0,
    PAGIdentitySelfAttnProcessor2_0,
2929
]