modeling_utils.py 49.7 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 os
20
import re
21
from collections import OrderedDict
22
from functools import partial
23
from pathlib import Path
24
from typing import Any, Callable, List, Optional, Tuple, Union
25

26
import safetensors
27
import torch
28
from huggingface_hub import create_repo
29
from huggingface_hub.utils import validate_hf_hub_args
30
from torch import Tensor, nn
31

32
33
from .. import __version__
from ..utils import (
34
    CONFIG_NAME,
35
    FLAX_WEIGHTS_NAME,
36
    SAFETENSORS_WEIGHTS_NAME,
37
    WEIGHTS_NAME,
38
39
    _add_variant,
    _get_model_file,
40
    deprecate,
41
42
43
44
    is_accelerate_available,
    is_torch_version,
    logging,
)
45
46
47
48
49
from ..utils.hub_utils import (
    PushToHubMixin,
    load_or_create_model_card,
    populate_model_card,
)
50
51
52
53
54
55
from .model_loading_utils import (
    _determine_device_map,
    _load_state_dict_into_model,
    load_model_dict_into_meta,
    load_state_dict,
)
56
57
58
59
60


logger = logging.get_logger(__name__)


61
62
63
64
65
66
if is_torch_version(">=", "1.9.0"):
    _LOW_CPU_MEM_USAGE_DEFAULT = True
else:
    _LOW_CPU_MEM_USAGE_DEFAULT = False


67
68
69
70
if is_accelerate_available():
    import accelerate


71
def get_parameter_device(parameter: torch.nn.Module) -> torch.device:
72
    try:
Patrick von Platen's avatar
Patrick von Platen committed
73
74
        parameters_and_buffers = itertools.chain(parameter.parameters(), parameter.buffers())
        return next(parameters_and_buffers).device
75
76
77
78
79
80
81
82
83
84
85
86
    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


87
def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype:
88
    try:
89
90
91
92
93
94
95
96
        params = tuple(parameter.parameters())
        if len(params) > 0:
            return params[0].dtype

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

97
98
99
100
101
102
103
104
105
106
107
108
    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


109
class ModelMixin(torch.nn.Module, PushToHubMixin):
110
111
112
    r"""
    Base class for all models.

Steven Liu's avatar
Steven Liu committed
113
114
    [`ModelMixin`] takes care of storing the model configuration and provides methods for loading, downloading and
    saving models.
115

Steven Liu's avatar
Steven Liu committed
116
        - **config_name** ([`str`]) -- Filename to save a model to when calling [`~models.ModelMixin.save_pretrained`].
117
    """
118

119
    config_name = CONFIG_NAME
Patrick von Platen's avatar
Patrick von Platen committed
120
    _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]
121
    _supports_gradient_checkpointing = False
122
    _keys_to_ignore_on_load_unexpected = None
123
    _no_split_modules = None
124

125
    def __init__(self):
126
127
        super().__init__()

128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
    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)

146
147
148
149
150
151
152
    @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())

153
    def enable_gradient_checkpointing(self) -> None:
154
        """
Steven Liu's avatar
Steven Liu committed
155
156
        Activates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
        *checkpoint activations* in other frameworks).
157
158
159
160
161
        """
        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))

162
    def disable_gradient_checkpointing(self) -> None:
163
        """
Steven Liu's avatar
Steven Liu committed
164
165
        Deactivates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
        *checkpoint activations* in other frameworks).
166
167
168
169
        """
        if self._supports_gradient_checkpointing:
            self.apply(partial(self._set_gradient_checkpointing, value=False))

170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
    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)

200
201
202
    def set_use_memory_efficient_attention_xformers(
        self, valid: bool, attention_op: Optional[Callable] = None
    ) -> None:
203
204
205
206
207
        # 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"):
208
                module.set_use_memory_efficient_attention_xformers(valid, attention_op)
209
210
211
212
213
214
215
216

            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)

217
    def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None) -> None:
218
        r"""
Steven Liu's avatar
Steven Liu committed
219
        Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
220

Steven Liu's avatar
Steven Liu committed
221
222
        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.
223

Steven Liu's avatar
Steven Liu committed
224
225
226
227
228
229
        <Tip warning={true}>

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

        </Tip>
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249

        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)
        ```
250
        """
251
        self.set_use_memory_efficient_attention_xformers(True, attention_op)
252

253
    def disable_xformers_memory_efficient_attention(self) -> None:
254
        r"""
Steven Liu's avatar
Steven Liu committed
255
        Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
256
257
258
        """
        self.set_use_memory_efficient_attention_xformers(False)

259
260
261
262
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
        is_main_process: bool = True,
263
        save_function: Optional[Callable] = None,
264
        safe_serialization: bool = True,
265
        variant: Optional[str] = None,
266
267
        push_to_hub: bool = False,
        **kwargs,
268
269
    ):
        """
Steven Liu's avatar
Steven Liu committed
270
271
        Save a model and its configuration file to a directory so that it can be reloaded using the
        [`~models.ModelMixin.from_pretrained`] class method.
272
273
274

        Arguments:
            save_directory (`str` or `os.PathLike`):
Steven Liu's avatar
Steven Liu committed
275
                Directory to save a model and its configuration file to. Will be created if it doesn't exist.
276
            is_main_process (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
277
278
279
                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.
280
            save_function (`Callable`):
Steven Liu's avatar
Steven Liu committed
281
282
                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
283
                `DIFFUSERS_SAVE_MODE`.
284
            safe_serialization (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
285
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
286
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
287
                If specified, weights are saved in the format `pytorch_model.<variant>.bin`.
288
289
290
291
292
293
            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.
294
295
296
297
298
299
300
        """
        if os.path.isfile(save_directory):
            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
            return

        os.makedirs(save_directory, exist_ok=True)

301
302
303
304
305
306
307
308
309
        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
310
311
312
313
314
        model_to_save = self

        # Attach architecture to the config
        # Save the config
        if is_main_process:
315
            model_to_save.save_config(save_directory)
316
317
318
319

        # Save the model
        state_dict = model_to_save.state_dict()

320
        weights_name = SAFETENSORS_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
321
        weights_name = _add_variant(weights_name, variant)
322

323
        # Save the model
324
325
        if safe_serialization:
            safetensors.torch.save_file(
326
                state_dict, Path(save_directory, weights_name).as_posix(), metadata={"format": "pt"}
327
328
            )
        else:
329
            torch.save(state_dict, Path(save_directory, weights_name).as_posix())
330

331
        logger.info(f"Model weights saved in {Path(save_directory, weights_name).as_posix()}")
332

333
        if push_to_hub:
334
335
336
            # 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)
337
            model_card.save(Path(save_directory, "README.md").as_posix())
338

339
340
341
342
343
344
345
346
            self._upload_folder(
                save_directory,
                repo_id,
                token=token,
                commit_message=commit_message,
                create_pr=create_pr,
            )

347
    @classmethod
348
    @validate_hf_hub_args
349
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
350
        r"""
Steven Liu's avatar
Steven Liu committed
351
        Instantiate a pretrained PyTorch model from a pretrained model configuration.
352

Steven Liu's avatar
Steven Liu committed
353
354
        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()`.
355
356
357
358
359

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

Steven Liu's avatar
Steven Liu committed
360
361
362
363
                    - 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`].
364
365

            cache_dir (`Union[str, os.PathLike]`, *optional*):
Steven Liu's avatar
Steven Liu committed
366
367
                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
368
            torch_dtype (`str` or `torch.dtype`, *optional*):
Steven Liu's avatar
Steven Liu committed
369
370
                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.
371
372
373
            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.
374
375
376
            resume_download:
                Deprecated and ignored. All downloads are now resumed by default when possible. Will be removed in v1
                of Diffusers.
377
            proxies (`Dict[str, str]`, *optional*):
Steven Liu's avatar
Steven Liu committed
378
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
379
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
Steven Liu's avatar
Steven Liu committed
380
            output_loading_info (`bool`, *optional*, defaults to `False`):
381
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
382
            local_files_only(`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
383
384
                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.
385
            token (`str` or *bool*, *optional*):
Steven Liu's avatar
Steven Liu committed
386
387
                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.
388
            revision (`str`, *optional*, defaults to `"main"`):
Steven Liu's avatar
Steven Liu committed
389
390
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
391
392
            from_flax (`bool`, *optional*, defaults to `False`):
                Load the model weights from a Flax checkpoint save file.
393
            subfolder (`str`, *optional*, defaults to `""`):
Steven Liu's avatar
Steven Liu committed
394
                The subfolder location of a model file within a larger model repository on the Hub or locally.
395
            mirror (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
396
397
398
                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.
399
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
Steven Liu's avatar
Steven Liu committed
400
401
                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
402
403
                same device.

Steven Liu's avatar
Steven Liu committed
404
                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
405
406
                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).
407
            max_memory (`Dict`, *optional*):
Steven Liu's avatar
Steven Liu committed
408
409
                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.
410
            offload_folder (`str` or `os.PathLike`, *optional*):
Steven Liu's avatar
Steven Liu committed
411
                The path to offload weights if `device_map` contains the value `"disk"`.
412
            offload_state_dict (`bool`, *optional*):
Steven Liu's avatar
Steven Liu committed
413
414
415
                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.
416
            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
Steven Liu's avatar
Steven Liu committed
417
418
419
420
                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.
421
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
422
423
                Load weights from a specified `variant` filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.
424
            use_safetensors (`bool`, *optional*, defaults to `None`):
Steven Liu's avatar
Steven Liu committed
425
426
427
                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.
428
429
430

        <Tip>

Steven Liu's avatar
Steven Liu committed
431
432
433
434
        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.
435
436
437

        </Tip>

Steven Liu's avatar
Steven Liu committed
438
        Example:
439

Steven Liu's avatar
Steven Liu committed
440
441
        ```py
        from diffusers import UNet2DConditionModel
442

Steven Liu's avatar
Steven Liu committed
443
444
445
446
        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:
447

Steven Liu's avatar
Steven Liu committed
448
449
450
451
452
        ```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.
        ```
453
        """
454
        cache_dir = kwargs.pop("cache_dir", None)
455
456
        ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
        force_download = kwargs.pop("force_download", False)
457
        from_flax = kwargs.pop("from_flax", False)
458
        resume_download = kwargs.pop("resume_download", None)
459
460
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
461
462
        local_files_only = kwargs.pop("local_files_only", None)
        token = kwargs.pop("token", None)
463
        revision = kwargs.pop("revision", None)
464
        torch_dtype = kwargs.pop("torch_dtype", None)
Patrick von Platen's avatar
Patrick von Platen committed
465
        subfolder = kwargs.pop("subfolder", None)
466
        device_map = kwargs.pop("device_map", None)
467
468
469
        max_memory = kwargs.pop("max_memory", None)
        offload_folder = kwargs.pop("offload_folder", None)
        offload_state_dict = kwargs.pop("offload_state_dict", False)
470
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
471
        variant = kwargs.pop("variant", None)
472
473
474
475
        use_safetensors = kwargs.pop("use_safetensors", None)

        allow_pickle = False
        if use_safetensors is None:
476
            use_safetensors = True
477
            allow_pickle = True
478

479
480
        if low_cpu_mem_usage and not is_accelerate_available():
            low_cpu_mem_usage = False
481
            logger.warning(
482
483
484
485
486
487
488
489
490
491
492
493
                "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`."
            )

494
495
        # 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"):
496
497
498
499
            raise NotImplementedError(
                "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
                " `device_map=None`."
            )
500

501
502
503
504
505
506
507
508
509
510
511
        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`."
            )
512

513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
        # 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.")

543
544
545
        # Load config if we don't provide a configuration
        config_path = pretrained_model_name_or_path

546
547
548
549
550
        user_agent = {
            "diffusers": __version__,
            "file_type": "model",
            "framework": "pytorch",
        }
551

552
553
554
555
556
557
558
559
560
561
        # 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,
562
            token=token,
563
564
565
566
567
568
569
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            **kwargs,
        )

        # load model
570
        model_file = None
571
        if from_flax:
572
            model_file = _get_model_file(
573
                pretrained_model_name_or_path,
574
                weights_name=FLAX_WEIGHTS_NAME,
575
576
577
578
579
                cache_dir=cache_dir,
                force_download=force_download,
                resume_download=resume_download,
                proxies=proxies,
                local_files_only=local_files_only,
580
                token=token,
581
582
583
                revision=revision,
                subfolder=subfolder,
                user_agent=user_agent,
584
                commit_hash=commit_hash,
585
586
            )
            model = cls.from_config(config, **unused_kwargs)
587

588
589
590
591
592
            # 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:
593
            if use_safetensors:
594
                try:
595
                    model_file = _get_model_file(
596
                        pretrained_model_name_or_path,
597
                        weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant),
598
599
600
601
602
                        cache_dir=cache_dir,
                        force_download=force_download,
                        resume_download=resume_download,
                        proxies=proxies,
                        local_files_only=local_files_only,
603
                        token=token,
604
605
606
                        revision=revision,
                        subfolder=subfolder,
                        user_agent=user_agent,
607
                        commit_hash=commit_hash,
608
                    )
609
610
611
                except IOError as e:
                    if not allow_pickle:
                        raise e
612
613
                    pass
            if model_file is None:
614
                model_file = _get_model_file(
615
                    pretrained_model_name_or_path,
616
                    weights_name=_add_variant(WEIGHTS_NAME, variant),
617
618
619
620
621
                    cache_dir=cache_dir,
                    force_download=force_download,
                    resume_download=resume_download,
                    proxies=proxies,
                    local_files_only=local_files_only,
622
                    token=token,
623
624
625
                    revision=revision,
                    subfolder=subfolder,
                    user_agent=user_agent,
626
                    commit_hash=commit_hash,
627
628
629
630
631
632
633
634
635
636
                )

            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
                if device_map is None:
                    param_device = "cpu"
637
                    state_dict = load_state_dict(model_file, variant=variant)
638
                    model._convert_deprecated_attention_blocks(state_dict)
639
                    # move the params from meta device to cpu
640
641
642
643
644
                    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
645
                            " `low_cpu_mem_usage=False` and `device_map=None` if you want to randomly initialize"
646
647
                            " those weights or else make sure your checkpoint file is correct."
                        )
648

649
650
651
652
653
654
655
                    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,
                    )
656
657
658
659
660
661

                    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:
662
                        logger.warning(
663
664
665
                            f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
                        )

666
667
                else:  # else let accelerate handle loading and dispatching.
                    # Load weights and dispatch according to the device_map
Alexander Pivovarov's avatar
Alexander Pivovarov committed
668
                    # by default the device_map is None and the weights are loaded on the CPU
669
                    device_map = _determine_device_map(model, device_map, max_memory, torch_dtype)
670
671
672
673
674
675
676
677
678
                    try:
                        accelerate.load_checkpoint_and_dispatch(
                            model,
                            model_file,
                            device_map,
                            max_memory=max_memory,
                            offload_folder=offload_folder,
                            offload_state_dict=offload_state_dict,
                            dtype=torch_dtype,
679
                            force_hooks=True,
680
                            strict=True,
681
682
683
684
685
686
687
688
689
690
691
692
                        )
                    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):
693
                            logger.warning(
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
                                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,
                                model_file,
                                device_map,
                                max_memory=max_memory,
                                offload_folder=offload_folder,
                                offload_state_dict=offload_state_dict,
                                dtype=torch_dtype,
                            )
                            model._undo_temp_convert_self_to_deprecated_attention_blocks()
                        else:
                            raise e
713
714
715
716
717
718
719
720

                loading_info = {
                    "missing_keys": [],
                    "unexpected_keys": [],
                    "mismatched_keys": [],
                    "error_msgs": [],
                }
            else:
721
                model = cls.from_config(config, **unused_kwargs)
722

723
                state_dict = load_state_dict(model_file, variant=variant)
724
                model._convert_deprecated_attention_blocks(state_dict)
725

726
727
728
729
730
731
732
                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,
                )
733

734
735
736
737
738
739
                loading_info = {
                    "missing_keys": missing_keys,
                    "unexpected_keys": unexpected_keys,
                    "mismatched_keys": mismatched_keys,
                    "error_msgs": error_msgs,
                }
740
741
742
743
744
745
746
747
748
749
750
751
752

        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:
753
754
755
756
757
758
759
760
            return model, loading_info

        return model

    @classmethod
    def _load_pretrained_model(
        cls,
        model,
761
        state_dict: OrderedDict,
762
        resolved_archive_file,
763
764
        pretrained_model_name_or_path: Union[str, os.PathLike],
        ignore_mismatched_sizes: bool = False,
765
766
767
    ):
        # Retrieve missing & unexpected_keys
        model_state_dict = model.state_dict()
768
        loaded_keys = list(state_dict.keys())
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818

        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
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
        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."
            )
858
859
860

        return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs

861
862
863
864
865
866
867
868
869
    @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

870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
    # 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)

900
    @property
901
    def device(self) -> torch.device:
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
        """
        `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
917
        Get number of (trainable or non-embedding) parameters in the module.
918
919
920

        Args:
            only_trainable (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
921
                Whether or not to return only the number of trainable parameters.
922
            exclude_embeddings (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
923
                Whether or not to return only the number of non-embedding parameters.
924
925
926

        Returns:
            `int`: The number of parameters.
Steven Liu's avatar
Steven Liu committed
927
928
929
930
931
932
933
934
935
936
937

        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
        ```
938
939
940
941
942
943
944
945
946
947
948
949
950
951
        """

        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)
952

953
    def _convert_deprecated_attention_blocks(self, state_dict: OrderedDict) -> None:
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
        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")
996

997
    def _temp_convert_self_to_deprecated_attention_blocks(self) -> None:
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
        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

1024
    def _undo_temp_convert_self_to_deprecated_attention_blocks(self) -> None:
1025
1026
        deprecated_attention_block_modules = []

1027
        def recursive_find_attn_block(module) -> None:
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
            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
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


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

        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)

        return remapped_class.from_pretrained(pretrained_model_name_or_path, **kwargs)