modeling_utils.py 55.9 KB
Newer Older
1
# coding=utf-8
2
# Copyright 2024 The HuggingFace Inc. team.
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

17
import inspect
Patrick von Platen's avatar
Patrick von Platen committed
18
import itertools
19
import json
20
import os
21
import re
22
from collections import OrderedDict
23
from functools import partial
24
from pathlib import Path
25
from typing import Any, Callable, List, Optional, Tuple, Union
26

27
import safetensors
28
import torch
29
from huggingface_hub import create_repo, split_torch_state_dict_into_shards
30
from huggingface_hub.utils import validate_hf_hub_args
31
from torch import Tensor, nn
32

33
34
from .. import __version__
from ..utils import (
35
    CONFIG_NAME,
36
    FLAX_WEIGHTS_NAME,
37
    SAFE_WEIGHTS_INDEX_NAME,
38
    SAFETENSORS_WEIGHTS_NAME,
39
    WEIGHTS_INDEX_NAME,
40
    WEIGHTS_NAME,
41
    _add_variant,
42
    _get_checkpoint_shard_files,
43
    _get_model_file,
44
    deprecate,
45
46
47
48
    is_accelerate_available,
    is_torch_version,
    logging,
)
49
50
51
52
53
from ..utils.hub_utils import (
    PushToHubMixin,
    load_or_create_model_card,
    populate_model_card,
)
54
55
from .model_loading_utils import (
    _determine_device_map,
56
    _fetch_index_file,
57
58
59
60
    _load_state_dict_into_model,
    load_model_dict_into_meta,
    load_state_dict,
)
61
62
63
64


logger = logging.get_logger(__name__)

65
66
_REGEX_SHARD = re.compile(r"(.*?)-\d{5}-of-\d{5}")

67

68
69
70
71
72
73
if is_torch_version(">=", "1.9.0"):
    _LOW_CPU_MEM_USAGE_DEFAULT = True
else:
    _LOW_CPU_MEM_USAGE_DEFAULT = False


74
75
76
77
if is_accelerate_available():
    import accelerate


78
def get_parameter_device(parameter: torch.nn.Module) -> torch.device:
79
    try:
Patrick von Platen's avatar
Patrick von Platen committed
80
81
        parameters_and_buffers = itertools.chain(parameter.parameters(), parameter.buffers())
        return next(parameters_and_buffers).device
82
83
84
85
86
87
88
89
90
91
92
93
    except StopIteration:
        # For torch.nn.DataParallel compatibility in PyTorch 1.5

        def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]:
            tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
            return tuples

        gen = parameter._named_members(get_members_fn=find_tensor_attributes)
        first_tuple = next(gen)
        return first_tuple[1].device


94
def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype:
95
    try:
96
97
98
99
100
101
102
103
        params = tuple(parameter.parameters())
        if len(params) > 0:
            return params[0].dtype

        buffers = tuple(parameter.buffers())
        if len(buffers) > 0:
            return buffers[0].dtype

104
105
106
107
108
109
110
111
112
113
114
115
    except StopIteration:
        # For torch.nn.DataParallel compatibility in PyTorch 1.5

        def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]:
            tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
            return tuples

        gen = parameter._named_members(get_members_fn=find_tensor_attributes)
        first_tuple = next(gen)
        return first_tuple[1].dtype


116
class ModelMixin(torch.nn.Module, PushToHubMixin):
117
118
119
    r"""
    Base class for all models.

Steven Liu's avatar
Steven Liu committed
120
121
    [`ModelMixin`] takes care of storing the model configuration and provides methods for loading, downloading and
    saving models.
122

Steven Liu's avatar
Steven Liu committed
123
        - **config_name** ([`str`]) -- Filename to save a model to when calling [`~models.ModelMixin.save_pretrained`].
124
    """
125

126
    config_name = CONFIG_NAME
Patrick von Platen's avatar
Patrick von Platen committed
127
    _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]
128
    _supports_gradient_checkpointing = False
129
    _keys_to_ignore_on_load_unexpected = None
130
    _no_split_modules = None
131

132
    def __init__(self):
133
134
        super().__init__()

135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
    def __getattr__(self, name: str) -> Any:
        """The only reason we overwrite `getattr` here is to gracefully deprecate accessing
        config attributes directly. See https://github.com/huggingface/diffusers/pull/3129 We need to overwrite
        __getattr__ here in addition so that we don't trigger `torch.nn.Module`'s __getattr__':
        https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
        """

        is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name)
        is_attribute = name in self.__dict__

        if is_in_config and not is_attribute:
            deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'unet.config.{name}'."
            deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False, stacklevel=3)
            return self._internal_dict[name]

        # call PyTorch's https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
        return super().__getattr__(name)

153
154
155
156
157
158
159
    @property
    def is_gradient_checkpointing(self) -> bool:
        """
        Whether gradient checkpointing is activated for this model or not.
        """
        return any(hasattr(m, "gradient_checkpointing") and m.gradient_checkpointing for m in self.modules())

160
    def enable_gradient_checkpointing(self) -> None:
161
        """
Steven Liu's avatar
Steven Liu committed
162
163
        Activates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
        *checkpoint activations* in other frameworks).
164
165
166
167
168
        """
        if not self._supports_gradient_checkpointing:
            raise ValueError(f"{self.__class__.__name__} does not support gradient checkpointing.")
        self.apply(partial(self._set_gradient_checkpointing, value=True))

169
    def disable_gradient_checkpointing(self) -> None:
170
        """
Steven Liu's avatar
Steven Liu committed
171
172
        Deactivates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
        *checkpoint activations* in other frameworks).
173
174
175
176
        """
        if self._supports_gradient_checkpointing:
            self.apply(partial(self._set_gradient_checkpointing, value=False))

177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
    def set_use_npu_flash_attention(self, valid: bool) -> None:
        r"""
        Set the switch for the npu flash attention.
        """

        def fn_recursive_set_npu_flash_attention(module: torch.nn.Module):
            if hasattr(module, "set_use_npu_flash_attention"):
                module.set_use_npu_flash_attention(valid)

            for child in module.children():
                fn_recursive_set_npu_flash_attention(child)

        for module in self.children():
            if isinstance(module, torch.nn.Module):
                fn_recursive_set_npu_flash_attention(module)

    def enable_npu_flash_attention(self) -> None:
        r"""
        Enable npu flash attention from torch_npu

        """
        self.set_use_npu_flash_attention(True)

    def disable_npu_flash_attention(self) -> None:
        r"""
        disable npu flash attention from torch_npu

        """
        self.set_use_npu_flash_attention(False)

207
208
209
    def set_use_memory_efficient_attention_xformers(
        self, valid: bool, attention_op: Optional[Callable] = None
    ) -> None:
210
211
212
213
214
        # Recursively walk through all the children.
        # Any children which exposes the set_use_memory_efficient_attention_xformers method
        # gets the message
        def fn_recursive_set_mem_eff(module: torch.nn.Module):
            if hasattr(module, "set_use_memory_efficient_attention_xformers"):
215
                module.set_use_memory_efficient_attention_xformers(valid, attention_op)
216
217
218
219
220
221
222
223

            for child in module.children():
                fn_recursive_set_mem_eff(child)

        for module in self.children():
            if isinstance(module, torch.nn.Module):
                fn_recursive_set_mem_eff(module)

224
    def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None) -> None:
225
        r"""
Steven Liu's avatar
Steven Liu committed
226
        Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
227

Steven Liu's avatar
Steven Liu committed
228
229
        When this option is enabled, you should observe lower GPU memory usage and a potential speed up during
        inference. Speed up during training is not guaranteed.
230

Steven Liu's avatar
Steven Liu committed
231
232
233
234
235
236
        <Tip warning={true}>

        ⚠️ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes
        precedent.

        </Tip>
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256

        Parameters:
            attention_op (`Callable`, *optional*):
                Override the default `None` operator for use as `op` argument to the
                [`memory_efficient_attention()`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention)
                function of xFormers.

        Examples:

        ```py
        >>> import torch
        >>> from diffusers import UNet2DConditionModel
        >>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp

        >>> model = UNet2DConditionModel.from_pretrained(
        ...     "stabilityai/stable-diffusion-2-1", subfolder="unet", torch_dtype=torch.float16
        ... )
        >>> model = model.to("cuda")
        >>> model.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
        ```
257
        """
258
        self.set_use_memory_efficient_attention_xformers(True, attention_op)
259

260
    def disable_xformers_memory_efficient_attention(self) -> None:
261
        r"""
Steven Liu's avatar
Steven Liu committed
262
        Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
263
264
265
        """
        self.set_use_memory_efficient_attention_xformers(False)

266
267
268
269
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
        is_main_process: bool = True,
270
        save_function: Optional[Callable] = None,
271
        safe_serialization: bool = True,
272
        variant: Optional[str] = None,
273
        max_shard_size: Union[int, str] = "10GB",
274
275
        push_to_hub: bool = False,
        **kwargs,
276
277
    ):
        """
Steven Liu's avatar
Steven Liu committed
278
279
        Save a model and its configuration file to a directory so that it can be reloaded using the
        [`~models.ModelMixin.from_pretrained`] class method.
280
281
282

        Arguments:
            save_directory (`str` or `os.PathLike`):
Steven Liu's avatar
Steven Liu committed
283
                Directory to save a model and its configuration file to. Will be created if it doesn't exist.
284
            is_main_process (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
285
286
287
                Whether the process calling this is the main process or not. Useful during distributed training and you
                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
                process to avoid race conditions.
288
            save_function (`Callable`):
Steven Liu's avatar
Steven Liu committed
289
290
                The function to use to save the state dictionary. Useful during distributed training when you need to
                replace `torch.save` with another method. Can be configured with the environment variable
291
                `DIFFUSERS_SAVE_MODE`.
292
            safe_serialization (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
293
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
294
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
295
                If specified, weights are saved in the format `pytorch_model.<variant>.bin`.
296
            max_shard_size (`int` or `str`, defaults to `"10GB"`):
297
298
                The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size
                lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5GB"`).
299
300
301
302
                If expressed as an integer, the unit is bytes. Note that this limit will be decreased after a certain
                period of time (starting from Oct 2024) to allow users to upgrade to the latest version of `diffusers`.
                This is to establish a common default size for this argument across different libraries in the Hugging
                Face ecosystem (`transformers`, and `accelerate`, for example).
303
304
305
306
307
308
            push_to_hub (`bool`, *optional*, defaults to `False`):
                Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the
                repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
                namespace).
            kwargs (`Dict[str, Any]`, *optional*):
                Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
309
310
311
312
313
        """
        if os.path.isfile(save_directory):
            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
            return

314
315
316
317
318
319
320
321
        weights_name = SAFETENSORS_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
        weights_name = _add_variant(weights_name, variant)
        weight_name_split = weights_name.split(".")
        if len(weight_name_split) in [2, 3]:
            weights_name_pattern = weight_name_split[0] + "{suffix}." + ".".join(weight_name_split[1:])
        else:
            raise ValueError(f"Invalid {weights_name} provided.")

322
323
        os.makedirs(save_directory, exist_ok=True)

324
325
326
327
328
329
330
331
332
        if push_to_hub:
            commit_message = kwargs.pop("commit_message", None)
            private = kwargs.pop("private", False)
            create_pr = kwargs.pop("create_pr", False)
            token = kwargs.pop("token", None)
            repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
            repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id

        # Only save the model itself if we are using distributed training
333
334
335
336
337
        model_to_save = self

        # Attach architecture to the config
        # Save the config
        if is_main_process:
338
            model_to_save.save_config(save_directory)
339
340
341
342
343

        # Save the model
        state_dict = model_to_save.state_dict()

        # Save the model
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
        state_dict_split = split_torch_state_dict_into_shards(
            state_dict, max_shard_size=max_shard_size, filename_pattern=weights_name_pattern
        )

        # Clean the folder from a previous save
        if is_main_process:
            for filename in os.listdir(save_directory):
                if filename in state_dict_split.filename_to_tensors.keys():
                    continue
                full_filename = os.path.join(save_directory, filename)
                if not os.path.isfile(full_filename):
                    continue
                weights_without_ext = weights_name_pattern.replace(".bin", "").replace(".safetensors", "")
                weights_without_ext = weights_without_ext.replace("{suffix}", "")
                filename_without_ext = filename.replace(".bin", "").replace(".safetensors", "")
                # make sure that file to be deleted matches format of sharded file, e.g. pytorch_model-00001-of-00005
                if (
                    filename.startswith(weights_without_ext)
                    and _REGEX_SHARD.fullmatch(filename_without_ext) is not None
                ):
                    os.remove(full_filename)

        for filename, tensors in state_dict_split.filename_to_tensors.items():
            shard = {tensor: state_dict[tensor] for tensor in tensors}
            filepath = os.path.join(save_directory, filename)
            if safe_serialization:
                # At some point we will need to deal better with save_function (used for TPU and other distributed
                # joyfulness), but for now this enough.
                safetensors.torch.save_file(shard, filepath, metadata={"format": "pt"})
            else:
                torch.save(shard, filepath)

        if state_dict_split.is_sharded:
            index = {
                "metadata": state_dict_split.metadata,
                "weight_map": state_dict_split.tensor_to_filename,
            }
            save_index_file = SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
            save_index_file = os.path.join(save_directory, _add_variant(save_index_file, variant))
            # Save the index as well
            with open(save_index_file, "w", encoding="utf-8") as f:
                content = json.dumps(index, indent=2, sort_keys=True) + "\n"
                f.write(content)
            logger.info(
                f"The model is bigger than the maximum size per checkpoint ({max_shard_size}) and is going to be "
                f"split in {len(state_dict_split.filename_to_tensors)} checkpoint shards. You can find where each parameters has been saved in the "
                f"index located at {save_index_file}."
391
392
            )
        else:
393
394
            path_to_weights = os.path.join(save_directory, weights_name)
            logger.info(f"Model weights saved in {path_to_weights}")
395

396
        if push_to_hub:
397
398
399
            # Create a new empty model card and eventually tag it
            model_card = load_or_create_model_card(repo_id, token=token)
            model_card = populate_model_card(model_card)
400
            model_card.save(Path(save_directory, "README.md").as_posix())
401

402
403
404
405
406
407
408
409
            self._upload_folder(
                save_directory,
                repo_id,
                token=token,
                commit_message=commit_message,
                create_pr=create_pr,
            )

410
    @classmethod
411
    @validate_hf_hub_args
412
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
413
        r"""
Steven Liu's avatar
Steven Liu committed
414
        Instantiate a pretrained PyTorch model from a pretrained model configuration.
415

Steven Liu's avatar
Steven Liu committed
416
417
        The model is set in evaluation mode - `model.eval()` - by default, and dropout modules are deactivated. To
        train the model, set it back in training mode with `model.train()`.
418
419
420
421
422

        Parameters:
            pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
                Can be either:

Steven Liu's avatar
Steven Liu committed
423
424
425
426
                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
                      the Hub.
                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
                      with [`~ModelMixin.save_pretrained`].
427
428

            cache_dir (`Union[str, os.PathLike]`, *optional*):
Steven Liu's avatar
Steven Liu committed
429
430
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
Kashif Rasul's avatar
Kashif Rasul committed
431
            torch_dtype (`str` or `torch.dtype`, *optional*):
Steven Liu's avatar
Steven Liu committed
432
433
                Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
                dtype is automatically derived from the model's weights.
434
435
436
            force_download (`bool`, *optional*, defaults to `False`):
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
437
438
439
            resume_download:
                Deprecated and ignored. All downloads are now resumed by default when possible. Will be removed in v1
                of Diffusers.
440
            proxies (`Dict[str, str]`, *optional*):
Steven Liu's avatar
Steven Liu committed
441
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
442
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
Steven Liu's avatar
Steven Liu committed
443
            output_loading_info (`bool`, *optional*, defaults to `False`):
444
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
445
            local_files_only(`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
446
447
                Whether to only load local model weights and configuration files or not. If set to `True`, the model
                won't be downloaded from the Hub.
448
            token (`str` or *bool*, *optional*):
Steven Liu's avatar
Steven Liu committed
449
450
                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                `diffusers-cli login` (stored in `~/.huggingface`) is used.
451
            revision (`str`, *optional*, defaults to `"main"`):
Steven Liu's avatar
Steven Liu committed
452
453
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
454
455
            from_flax (`bool`, *optional*, defaults to `False`):
                Load the model weights from a Flax checkpoint save file.
456
            subfolder (`str`, *optional*, defaults to `""`):
Steven Liu's avatar
Steven Liu committed
457
                The subfolder location of a model file within a larger model repository on the Hub or locally.
458
            mirror (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
459
460
461
                Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                information.
462
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
Steven Liu's avatar
Steven Liu committed
463
464
                A map that specifies where each submodule should go. It doesn't need to be defined for each
                parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
465
                same device. Defaults to `None`, meaning that the model will be loaded on CPU.
466

Steven Liu's avatar
Steven Liu committed
467
                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
468
469
                more information about each option see [designing a device
                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
470
            max_memory (`Dict`, *optional*):
Steven Liu's avatar
Steven Liu committed
471
472
                A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
                each GPU and the available CPU RAM if unset.
473
            offload_folder (`str` or `os.PathLike`, *optional*):
Steven Liu's avatar
Steven Liu committed
474
                The path to offload weights if `device_map` contains the value `"disk"`.
475
            offload_state_dict (`bool`, *optional*):
Steven Liu's avatar
Steven Liu committed
476
477
478
                If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
                the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
                when there is some disk offload.
479
            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
Steven Liu's avatar
Steven Liu committed
480
481
482
483
                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
                argument to `True` will raise an error.
484
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
485
486
                Load weights from a specified `variant` filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.
487
            use_safetensors (`bool`, *optional*, defaults to `None`):
Steven Liu's avatar
Steven Liu committed
488
489
490
                If set to `None`, the `safetensors` weights are downloaded if they're available **and** if the
                `safetensors` library is installed. If set to `True`, the model is forcibly loaded from `safetensors`
                weights. If set to `False`, `safetensors` weights are not loaded.
491
492
493

        <Tip>

Steven Liu's avatar
Steven Liu committed
494
495
496
497
        To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with
        `huggingface-cli login`. You can also activate the special
        ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use this method in a
        firewalled environment.
498
499
500

        </Tip>

Steven Liu's avatar
Steven Liu committed
501
        Example:
502

Steven Liu's avatar
Steven Liu committed
503
504
        ```py
        from diffusers import UNet2DConditionModel
505

Steven Liu's avatar
Steven Liu committed
506
507
508
509
        unet = UNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="unet")
        ```

        If you get the error message below, you need to finetune the weights for your downstream task:
510

Steven Liu's avatar
Steven Liu committed
511
512
513
514
515
        ```bash
        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
        - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
        You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
        ```
516
        """
517
        cache_dir = kwargs.pop("cache_dir", None)
518
519
        ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
        force_download = kwargs.pop("force_download", False)
520
        from_flax = kwargs.pop("from_flax", False)
521
        resume_download = kwargs.pop("resume_download", None)
522
523
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
524
525
        local_files_only = kwargs.pop("local_files_only", None)
        token = kwargs.pop("token", None)
526
        revision = kwargs.pop("revision", None)
527
        torch_dtype = kwargs.pop("torch_dtype", None)
Patrick von Platen's avatar
Patrick von Platen committed
528
        subfolder = kwargs.pop("subfolder", None)
529
        device_map = kwargs.pop("device_map", None)
530
531
532
        max_memory = kwargs.pop("max_memory", None)
        offload_folder = kwargs.pop("offload_folder", None)
        offload_state_dict = kwargs.pop("offload_state_dict", False)
533
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
534
        variant = kwargs.pop("variant", None)
535
536
537
538
        use_safetensors = kwargs.pop("use_safetensors", None)

        allow_pickle = False
        if use_safetensors is None:
539
            use_safetensors = True
540
            allow_pickle = True
541

542
543
        if low_cpu_mem_usage and not is_accelerate_available():
            low_cpu_mem_usage = False
544
            logger.warning(
545
546
547
548
549
550
551
552
553
554
555
556
                "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
                " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
                " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
                " install accelerate\n```\n."
            )

        if device_map is not None and not is_accelerate_available():
            raise NotImplementedError(
                "Loading and dispatching requires `accelerate`. Please make sure to install accelerate or set"
                " `device_map=None`. You can install accelerate with `pip install accelerate`."
            )

557
558
        # Check if we can handle device_map and dispatching the weights
        if device_map is not None and not is_torch_version(">=", "1.9.0"):
559
560
561
562
            raise NotImplementedError(
                "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
                " `device_map=None`."
            )
563

564
565
566
567
568
569
570
571
572
573
574
        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
            raise NotImplementedError(
                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
                " `low_cpu_mem_usage=False`."
            )

        if low_cpu_mem_usage is False and device_map is not None:
            raise ValueError(
                f"You cannot set `low_cpu_mem_usage` to `False` while using device_map={device_map} for loading and"
                " dispatching. Please make sure to set `low_cpu_mem_usage=True`."
            )
575

576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
        # change device_map into a map if we passed an int, a str or a torch.device
        if isinstance(device_map, torch.device):
            device_map = {"": device_map}
        elif isinstance(device_map, str) and device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]:
            try:
                device_map = {"": torch.device(device_map)}
            except RuntimeError:
                raise ValueError(
                    "When passing device_map as a string, the value needs to be a device name (e.g. cpu, cuda:0) or "
                    f"'auto', 'balanced', 'balanced_low_0', 'sequential' but found {device_map}."
                )
        elif isinstance(device_map, int):
            if device_map < 0:
                raise ValueError(
                    "You can't pass device_map as a negative int. If you want to put the model on the cpu, pass device_map = 'cpu' "
                )
            else:
                device_map = {"": device_map}

        if device_map is not None:
            if low_cpu_mem_usage is None:
                low_cpu_mem_usage = True
            elif not low_cpu_mem_usage:
                raise ValueError("Passing along a `device_map` requires `low_cpu_mem_usage=True`")

        if low_cpu_mem_usage:
            if device_map is not None and not is_torch_version(">=", "1.10"):
                # The max memory utils require PyTorch >= 1.10 to have torch.cuda.mem_get_info.
                raise ValueError("`low_cpu_mem_usage` and `device_map` require PyTorch >= 1.10.")

606
607
608
        # Load config if we don't provide a configuration
        config_path = pretrained_model_name_or_path

609
610
611
612
613
        user_agent = {
            "diffusers": __version__,
            "file_type": "model",
            "framework": "pytorch",
        }
614

615
616
617
618
619
620
621
622
623
624
        # load config
        config, unused_kwargs, commit_hash = cls.load_config(
            config_path,
            cache_dir=cache_dir,
            return_unused_kwargs=True,
            return_commit_hash=True,
            force_download=force_download,
            resume_download=resume_download,
            proxies=proxies,
            local_files_only=local_files_only,
625
            token=token,
626
627
628
629
630
631
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            **kwargs,
        )

632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
        # Determine if we're loading from a directory of sharded checkpoints.
        is_sharded = False
        index_file = None
        is_local = os.path.isdir(pretrained_model_name_or_path)
        index_file = _fetch_index_file(
            is_local=is_local,
            pretrained_model_name_or_path=pretrained_model_name_or_path,
            subfolder=subfolder or "",
            use_safetensors=use_safetensors,
            cache_dir=cache_dir,
            variant=variant,
            force_download=force_download,
            resume_download=resume_download,
            proxies=proxies,
            local_files_only=local_files_only,
            token=token,
            revision=revision,
            user_agent=user_agent,
            commit_hash=commit_hash,
        )
        if index_file is not None and index_file.is_file():
            is_sharded = True

        if is_sharded and from_flax:
            raise ValueError("Loading of sharded checkpoints is not supported when `from_flax=True`.")

658
        # load model
659
        model_file = None
660
        if from_flax:
661
            model_file = _get_model_file(
662
                pretrained_model_name_or_path,
663
                weights_name=FLAX_WEIGHTS_NAME,
664
665
666
667
668
                cache_dir=cache_dir,
                force_download=force_download,
                resume_download=resume_download,
                proxies=proxies,
                local_files_only=local_files_only,
669
                token=token,
670
671
672
                revision=revision,
                subfolder=subfolder,
                user_agent=user_agent,
673
                commit_hash=commit_hash,
674
675
            )
            model = cls.from_config(config, **unused_kwargs)
676

677
678
679
680
681
            # Convert the weights
            from .modeling_pytorch_flax_utils import load_flax_checkpoint_in_pytorch_model

            model = load_flax_checkpoint_in_pytorch_model(model, model_file)
        else:
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
            if is_sharded:
                sharded_ckpt_cached_folder, sharded_metadata = _get_checkpoint_shard_files(
                    pretrained_model_name_or_path,
                    index_file,
                    cache_dir=cache_dir,
                    proxies=proxies,
                    resume_download=resume_download,
                    local_files_only=local_files_only,
                    token=token,
                    user_agent=user_agent,
                    revision=revision,
                    subfolder=subfolder or "",
                )

            elif use_safetensors and not is_sharded:
697
                try:
698
                    model_file = _get_model_file(
699
                        pretrained_model_name_or_path,
700
                        weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant),
701
702
703
704
705
                        cache_dir=cache_dir,
                        force_download=force_download,
                        resume_download=resume_download,
                        proxies=proxies,
                        local_files_only=local_files_only,
706
                        token=token,
707
708
709
                        revision=revision,
                        subfolder=subfolder,
                        user_agent=user_agent,
710
                        commit_hash=commit_hash,
711
                    )
712

713
                except IOError as e:
714
                    logger.error(f"An error occurred while trying to fetch {pretrained_model_name_or_path}: {e}")
715
                    if not allow_pickle:
716
717
718
719
720
721
                        raise
                    logger.warning(
                        "Defaulting to unsafe serialization. Pass `allow_pickle=False` to raise an error instead."
                    )

            if model_file is None and not is_sharded:
722
                model_file = _get_model_file(
723
                    pretrained_model_name_or_path,
724
                    weights_name=_add_variant(WEIGHTS_NAME, variant),
725
726
727
728
729
                    cache_dir=cache_dir,
                    force_download=force_download,
                    resume_download=resume_download,
                    proxies=proxies,
                    local_files_only=local_files_only,
730
                    token=token,
731
732
733
                    revision=revision,
                    subfolder=subfolder,
                    user_agent=user_agent,
734
                    commit_hash=commit_hash,
735
736
737
738
739
740
741
742
                )

            if low_cpu_mem_usage:
                # Instantiate model with empty weights
                with accelerate.init_empty_weights():
                    model = cls.from_config(config, **unused_kwargs)

                # if device_map is None, load the state dict and move the params from meta device to the cpu
743
                if device_map is None and not is_sharded:
744
                    param_device = "cpu"
745
                    state_dict = load_state_dict(model_file, variant=variant)
746
                    model._convert_deprecated_attention_blocks(state_dict)
747
                    # move the params from meta device to cpu
748
749
750
751
752
                    missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())
                    if len(missing_keys) > 0:
                        raise ValueError(
                            f"Cannot load {cls} from {pretrained_model_name_or_path} because the following keys are"
                            f" missing: \n {', '.join(missing_keys)}. \n Please make sure to pass"
Alexander Pivovarov's avatar
Alexander Pivovarov committed
753
                            " `low_cpu_mem_usage=False` and `device_map=None` if you want to randomly initialize"
754
755
                            " those weights or else make sure your checkpoint file is correct."
                        )
756

757
758
759
760
761
762
763
                    unexpected_keys = load_model_dict_into_meta(
                        model,
                        state_dict,
                        device=param_device,
                        dtype=torch_dtype,
                        model_name_or_path=pretrained_model_name_or_path,
                    )
764
765
766
767
768
769

                    if cls._keys_to_ignore_on_load_unexpected is not None:
                        for pat in cls._keys_to_ignore_on_load_unexpected:
                            unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]

                    if len(unexpected_keys) > 0:
770
                        logger.warning(
771
772
773
                            f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
                        )

774
775
                else:  # else let accelerate handle loading and dispatching.
                    # Load weights and dispatch according to the device_map
Alexander Pivovarov's avatar
Alexander Pivovarov committed
776
                    # by default the device_map is None and the weights are loaded on the CPU
777
                    force_hook = True
778
                    device_map = _determine_device_map(model, device_map, max_memory, torch_dtype)
779
780
781
782
                    if device_map is None and is_sharded:
                        # we load the parameters on the cpu
                        device_map = {"": "cpu"}
                        force_hook = False
783
784
785
                    try:
                        accelerate.load_checkpoint_and_dispatch(
                            model,
786
                            model_file if not is_sharded else sharded_ckpt_cached_folder,
787
788
789
790
791
                            device_map,
                            max_memory=max_memory,
                            offload_folder=offload_folder,
                            offload_state_dict=offload_state_dict,
                            dtype=torch_dtype,
792
                            force_hooks=force_hook,
793
                            strict=True,
794
795
796
797
798
799
800
801
802
803
804
805
                        )
                    except AttributeError as e:
                        # When using accelerate loading, we do not have the ability to load the state
                        # dict and rename the weight names manually. Additionally, accelerate skips
                        # torch loading conventions and directly writes into `module.{_buffers, _parameters}`
                        # (which look like they should be private variables?), so we can't use the standard hooks
                        # to rename parameters on load. We need to mimic the original weight names so the correct
                        # attributes are available. After we have loaded the weights, we convert the deprecated
                        # names to the new non-deprecated names. Then we _greatly encourage_ the user to convert
                        # the weights so we don't have to do this again.

                        if "'Attention' object has no attribute" in str(e):
806
                            logger.warning(
807
808
809
810
811
812
813
814
815
                                f"Taking `{str(e)}` while using `accelerate.load_checkpoint_and_dispatch` to mean {pretrained_model_name_or_path}"
                                " was saved with deprecated attention block weight names. We will load it with the deprecated attention block"
                                " names and convert them on the fly to the new attention block format. Please re-save the model after this conversion,"
                                " so we don't have to do the on the fly renaming in the future. If the model is from a hub checkpoint,"
                                " please also re-upload it or open a PR on the original repository."
                            )
                            model._temp_convert_self_to_deprecated_attention_blocks()
                            accelerate.load_checkpoint_and_dispatch(
                                model,
816
                                model_file if not is_sharded else sharded_ckpt_cached_folder,
817
818
819
820
821
                                device_map,
                                max_memory=max_memory,
                                offload_folder=offload_folder,
                                offload_state_dict=offload_state_dict,
                                dtype=torch_dtype,
822
                                force_hooks=force_hook,
823
                                strict=True,
824
825
826
827
                            )
                            model._undo_temp_convert_self_to_deprecated_attention_blocks()
                        else:
                            raise e
828
829
830
831
832
833
834
835

                loading_info = {
                    "missing_keys": [],
                    "unexpected_keys": [],
                    "mismatched_keys": [],
                    "error_msgs": [],
                }
            else:
836
                model = cls.from_config(config, **unused_kwargs)
837

838
                state_dict = load_state_dict(model_file, variant=variant)
839
                model._convert_deprecated_attention_blocks(state_dict)
840

841
842
843
844
845
846
847
                model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model(
                    model,
                    state_dict,
                    model_file,
                    pretrained_model_name_or_path,
                    ignore_mismatched_sizes=ignore_mismatched_sizes,
                )
848

849
850
851
852
853
854
                loading_info = {
                    "missing_keys": missing_keys,
                    "unexpected_keys": unexpected_keys,
                    "mismatched_keys": mismatched_keys,
                    "error_msgs": error_msgs,
                }
855
856
857
858
859
860
861
862
863
864
865
866
867

        if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
            raise ValueError(
                f"{torch_dtype} needs to be of type `torch.dtype`, e.g. `torch.float16`, but is {type(torch_dtype)}."
            )
        elif torch_dtype is not None:
            model = model.to(torch_dtype)

        model.register_to_config(_name_or_path=pretrained_model_name_or_path)

        # Set model in evaluation mode to deactivate DropOut modules by default
        model.eval()
        if output_loading_info:
868
869
870
871
872
873
874
875
            return model, loading_info

        return model

    @classmethod
    def _load_pretrained_model(
        cls,
        model,
876
        state_dict: OrderedDict,
877
        resolved_archive_file,
878
879
        pretrained_model_name_or_path: Union[str, os.PathLike],
        ignore_mismatched_sizes: bool = False,
880
881
882
    ):
        # Retrieve missing & unexpected_keys
        model_state_dict = model.state_dict()
883
        loaded_keys = list(state_dict.keys())
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933

        expected_keys = list(model_state_dict.keys())

        original_loaded_keys = loaded_keys

        missing_keys = list(set(expected_keys) - set(loaded_keys))
        unexpected_keys = list(set(loaded_keys) - set(expected_keys))

        # Make sure we are able to load base models as well as derived models (with heads)
        model_to_load = model

        def _find_mismatched_keys(
            state_dict,
            model_state_dict,
            loaded_keys,
            ignore_mismatched_sizes,
        ):
            mismatched_keys = []
            if ignore_mismatched_sizes:
                for checkpoint_key in loaded_keys:
                    model_key = checkpoint_key

                    if (
                        model_key in model_state_dict
                        and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape
                    ):
                        mismatched_keys.append(
                            (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
                        )
                        del state_dict[checkpoint_key]
            return mismatched_keys

        if state_dict is not None:
            # Whole checkpoint
            mismatched_keys = _find_mismatched_keys(
                state_dict,
                model_state_dict,
                original_loaded_keys,
                ignore_mismatched_sizes,
            )
            error_msgs = _load_state_dict_into_model(model_to_load, state_dict)

        if len(error_msgs) > 0:
            error_msg = "\n\t".join(error_msgs)
            if "size mismatch" in error_msg:
                error_msg += (
                    "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
                )
            raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")

Patrick von Platen's avatar
Patrick von Platen committed
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
        if len(unexpected_keys) > 0:
            logger.warning(
                f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when"
                f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are"
                f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task"
                " or with another architecture (e.g. initializing a BertForSequenceClassification model from a"
                " BertForPreTraining model).\n- This IS NOT expected if you are initializing"
                f" {model.__class__.__name__} from the checkpoint of a model that you expect to be exactly"
                " identical (initializing a BertForSequenceClassification model from a"
                " BertForSequenceClassification model)."
            )
        else:
            logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
        if len(missing_keys) > 0:
            logger.warning(
                f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
                f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"
                " TRAIN this model on a down-stream task to be able to use it for predictions and inference."
            )
        elif len(mismatched_keys) == 0:
            logger.info(
                f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"
                f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the"
                f" checkpoint was trained on, you can already use {model.__class__.__name__} for predictions"
                " without further training."
            )
        if len(mismatched_keys) > 0:
            mismatched_warning = "\n".join(
                [
                    f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated"
                    for key, shape1, shape2 in mismatched_keys
                ]
            )
            logger.warning(
                f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
                f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not"
                f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be"
                " able to use it for predictions and inference."
            )
973
974
975

        return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs

976
977
978
979
980
981
982
983
984
    @classmethod
    def _get_signature_keys(cls, obj):
        parameters = inspect.signature(obj.__init__).parameters
        required_parameters = {k: v for k, v in parameters.items() if v.default == inspect._empty}
        optional_parameters = set({k for k, v in parameters.items() if v.default != inspect._empty})
        expected_modules = set(required_parameters.keys()) - {"self"}

        return expected_modules, optional_parameters

985
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
    # Adapted from `transformers` modeling_utils.py
    def _get_no_split_modules(self, device_map: str):
        """
        Get the modules of the model that should not be spit when using device_map. We iterate through the modules to
        get the underlying `_no_split_modules`.

        Args:
            device_map (`str`):
                The device map value. Options are ["auto", "balanced", "balanced_low_0", "sequential"]

        Returns:
            `List[str]`: List of modules that should not be split
        """
        _no_split_modules = set()
        modules_to_check = [self]
        while len(modules_to_check) > 0:
            module = modules_to_check.pop(-1)
            # if the module does not appear in _no_split_modules, we also check the children
            if module.__class__.__name__ not in _no_split_modules:
                if isinstance(module, ModelMixin):
                    if module._no_split_modules is None:
                        raise ValueError(
                            f"{module.__class__.__name__} does not support `device_map='{device_map}'`. To implement support, the model "
                            "class needs to implement the `_no_split_modules` attribute."
                        )
                    else:
                        _no_split_modules = _no_split_modules | set(module._no_split_modules)
                modules_to_check += list(module.children())
        return list(_no_split_modules)

1015
    @property
1016
    def device(self) -> torch.device:
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
        """
        `torch.device`: The device on which the module is (assuming that all the module parameters are on the same
        device).
        """
        return get_parameter_device(self)

    @property
    def dtype(self) -> torch.dtype:
        """
        `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).
        """
        return get_parameter_dtype(self)

    def num_parameters(self, only_trainable: bool = False, exclude_embeddings: bool = False) -> int:
        """
Steven Liu's avatar
Steven Liu committed
1032
        Get number of (trainable or non-embedding) parameters in the module.
1033
1034
1035

        Args:
            only_trainable (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
1036
                Whether or not to return only the number of trainable parameters.
1037
            exclude_embeddings (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
1038
                Whether or not to return only the number of non-embedding parameters.
1039
1040
1041

        Returns:
            `int`: The number of parameters.
Steven Liu's avatar
Steven Liu committed
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052

        Example:

        ```py
        from diffusers import UNet2DConditionModel

        model_id = "runwayml/stable-diffusion-v1-5"
        unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet")
        unet.num_parameters(only_trainable=True)
        859520964
        ```
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
        """

        if exclude_embeddings:
            embedding_param_names = [
                f"{name}.weight"
                for name, module_type in self.named_modules()
                if isinstance(module_type, torch.nn.Embedding)
            ]
            non_embedding_parameters = [
                parameter for name, parameter in self.named_parameters() if name not in embedding_param_names
            ]
            return sum(p.numel() for p in non_embedding_parameters if p.requires_grad or not only_trainable)
        else:
            return sum(p.numel() for p in self.parameters() if p.requires_grad or not only_trainable)
1067

1068
    def _convert_deprecated_attention_blocks(self, state_dict: OrderedDict) -> None:
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
        deprecated_attention_block_paths = []

        def recursive_find_attn_block(name, module):
            if hasattr(module, "_from_deprecated_attn_block") and module._from_deprecated_attn_block:
                deprecated_attention_block_paths.append(name)

            for sub_name, sub_module in module.named_children():
                sub_name = sub_name if name == "" else f"{name}.{sub_name}"
                recursive_find_attn_block(sub_name, sub_module)

        recursive_find_attn_block("", self)

        # NOTE: we have to check if the deprecated parameters are in the state dict
        # because it is possible we are loading from a state dict that was already
        # converted

        for path in deprecated_attention_block_paths:
            # group_norm path stays the same

            # query -> to_q
            if f"{path}.query.weight" in state_dict:
                state_dict[f"{path}.to_q.weight"] = state_dict.pop(f"{path}.query.weight")
            if f"{path}.query.bias" in state_dict:
                state_dict[f"{path}.to_q.bias"] = state_dict.pop(f"{path}.query.bias")

            # key -> to_k
            if f"{path}.key.weight" in state_dict:
                state_dict[f"{path}.to_k.weight"] = state_dict.pop(f"{path}.key.weight")
            if f"{path}.key.bias" in state_dict:
                state_dict[f"{path}.to_k.bias"] = state_dict.pop(f"{path}.key.bias")

            # value -> to_v
            if f"{path}.value.weight" in state_dict:
                state_dict[f"{path}.to_v.weight"] = state_dict.pop(f"{path}.value.weight")
            if f"{path}.value.bias" in state_dict:
                state_dict[f"{path}.to_v.bias"] = state_dict.pop(f"{path}.value.bias")

            # proj_attn -> to_out.0
            if f"{path}.proj_attn.weight" in state_dict:
                state_dict[f"{path}.to_out.0.weight"] = state_dict.pop(f"{path}.proj_attn.weight")
            if f"{path}.proj_attn.bias" in state_dict:
                state_dict[f"{path}.to_out.0.bias"] = state_dict.pop(f"{path}.proj_attn.bias")
1111

1112
    def _temp_convert_self_to_deprecated_attention_blocks(self) -> None:
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
        deprecated_attention_block_modules = []

        def recursive_find_attn_block(module):
            if hasattr(module, "_from_deprecated_attn_block") and module._from_deprecated_attn_block:
                deprecated_attention_block_modules.append(module)

            for sub_module in module.children():
                recursive_find_attn_block(sub_module)

        recursive_find_attn_block(self)

        for module in deprecated_attention_block_modules:
            module.query = module.to_q
            module.key = module.to_k
            module.value = module.to_v
            module.proj_attn = module.to_out[0]

            # We don't _have_ to delete the old attributes, but it's helpful to ensure
            # that _all_ the weights are loaded into the new attributes and we're not
            # making an incorrect assumption that this model should be converted when
            # it really shouldn't be.
            del module.to_q
            del module.to_k
            del module.to_v
            del module.to_out

1139
    def _undo_temp_convert_self_to_deprecated_attention_blocks(self) -> None:
1140
1141
        deprecated_attention_block_modules = []

1142
        def recursive_find_attn_block(module) -> None:
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
            if hasattr(module, "_from_deprecated_attn_block") and module._from_deprecated_attn_block:
                deprecated_attention_block_modules.append(module)

            for sub_module in module.children():
                recursive_find_attn_block(sub_module)

        recursive_find_attn_block(self)

        for module in deprecated_attention_block_modules:
            module.to_q = module.query
            module.to_k = module.key
            module.to_v = module.value
            module.to_out = nn.ModuleList([module.proj_attn, nn.Dropout(module.dropout)])

            del module.query
            del module.key
            del module.value
            del module.proj_attn
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174


class LegacyModelMixin(ModelMixin):
    r"""
    A subclass of `ModelMixin` to resolve class mapping from legacy classes (like `Transformer2DModel`) to more
    pipeline-specific classes (like `DiTTransformer2DModel`).
    """

    @classmethod
    @validate_hf_hub_args
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
        # To prevent depedency import problem.
        from .model_loading_utils import _fetch_remapped_cls_from_config

1175
1176
1177
        # Create a copy of the kwargs so that we don't mess with the keyword arguments in the downstream calls.
        kwargs_copy = kwargs.copy()

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
1211
1212
1213
1214
        cache_dir = kwargs.pop("cache_dir", None)
        force_download = kwargs.pop("force_download", False)
        resume_download = kwargs.pop("resume_download", None)
        proxies = kwargs.pop("proxies", None)
        local_files_only = kwargs.pop("local_files_only", None)
        token = kwargs.pop("token", None)
        revision = kwargs.pop("revision", None)
        subfolder = kwargs.pop("subfolder", None)

        # Load config if we don't provide a configuration
        config_path = pretrained_model_name_or_path

        user_agent = {
            "diffusers": __version__,
            "file_type": "model",
            "framework": "pytorch",
        }

        # load config
        config, _, _ = cls.load_config(
            config_path,
            cache_dir=cache_dir,
            return_unused_kwargs=True,
            return_commit_hash=True,
            force_download=force_download,
            resume_download=resume_download,
            proxies=proxies,
            local_files_only=local_files_only,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            **kwargs,
        )
        # resolve remapping
        remapped_class = _fetch_remapped_cls_from_config(config, cls)

1215
        return remapped_class.from_pretrained(pretrained_model_name_or_path, **kwargs_copy)