modeling_utils.py 77 KB
Newer Older
1
# coding=utf-8
2
# Copyright 2025 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 copy
18
import inspect
Patrick von Platen's avatar
Patrick von Platen committed
19
import itertools
20
import json
21
import os
22
import re
23
from collections import OrderedDict
24
from functools import wraps
25
from pathlib import Path
Aryan's avatar
Aryan committed
26
from typing import Any, Callable, Dict, List, Optional, Tuple, Type, Union
27

28
import safetensors
29
import torch
30
import torch.utils.checkpoint
Marc Sun's avatar
Marc Sun committed
31
from huggingface_hub import DDUFEntry, create_repo, split_torch_state_dict_into_shards
32
from huggingface_hub.utils import validate_hf_hub_args
33
from torch import Tensor, nn
34

35
from .. import __version__
Aryan's avatar
Aryan committed
36
from ..hooks import apply_layerwise_casting
37
38
from ..quantizers import DiffusersAutoQuantizer, DiffusersQuantizer
from ..quantizers.quantization_config import QuantizationMethod
39
from ..utils import (
40
    CONFIG_NAME,
41
    FLAX_WEIGHTS_NAME,
42
    SAFE_WEIGHTS_INDEX_NAME,
43
    SAFETENSORS_WEIGHTS_NAME,
44
    WEIGHTS_INDEX_NAME,
45
    WEIGHTS_NAME,
46
    _add_variant,
47
    _get_checkpoint_shard_files,
48
    _get_model_file,
49
    deprecate,
50
    is_accelerate_available,
51
52
    is_bitsandbytes_available,
    is_bitsandbytes_version,
Aryan's avatar
Aryan committed
53
    is_peft_available,
54
55
56
    is_torch_version,
    logging,
)
57
58
59
60
61
from ..utils.hub_utils import (
    PushToHubMixin,
    load_or_create_model_card,
    populate_model_card,
)
62
63
from .model_loading_utils import (
    _determine_device_map,
64
    _fetch_index_file,
65
    _fetch_index_file_legacy,
66
    _load_state_dict_into_model,
67
    _merge_sharded_checkpoints,
68
69
70
    load_model_dict_into_meta,
    load_state_dict,
)
71
72
73
74


logger = logging.get_logger(__name__)

75
76
_REGEX_SHARD = re.compile(r"(.*?)-\d{5}-of-\d{5}")

77

78
79
80
81
82
83
if is_torch_version(">=", "1.9.0"):
    _LOW_CPU_MEM_USAGE_DEFAULT = True
else:
    _LOW_CPU_MEM_USAGE_DEFAULT = False


84
85
86
87
if is_accelerate_available():
    import accelerate


88
def get_parameter_device(parameter: torch.nn.Module) -> torch.device:
89
    try:
Patrick von Platen's avatar
Patrick von Platen committed
90
91
        parameters_and_buffers = itertools.chain(parameter.parameters(), parameter.buffers())
        return next(parameters_and_buffers).device
92
93
94
95
96
97
98
99
100
101
102
103
    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


104
def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype:
105
106
107
    """
    Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.
    """
Aryan's avatar
Aryan committed
108
109
110
111
112
113
114
115
116
117
118
    # 1. Check if we have attached any dtype modifying hooks (eg. layerwise casting)
    if isinstance(parameter, nn.Module):
        for name, submodule in parameter.named_modules():
            if not hasattr(submodule, "_diffusers_hook"):
                continue
            registry = submodule._diffusers_hook
            hook = registry.get_hook("layerwise_casting")
            if hook is not None:
                return hook.compute_dtype

    # 2. If no dtype modifying hooks are attached, return the dtype of the first floating point parameter/buffer
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
    last_dtype = None
    for param in parameter.parameters():
        last_dtype = param.dtype
        if param.is_floating_point():
            return param.dtype

    for buffer in parameter.buffers():
        last_dtype = buffer.dtype
        if buffer.is_floating_point():
            return buffer.dtype

    if last_dtype is not None:
        # if no floating dtype was found return whatever the first dtype is
        return last_dtype

    # For nn.DataParallel compatibility in PyTorch > 1.5
    def find_tensor_attributes(module: 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)
    last_tuple = None
    for tuple in gen:
        last_tuple = tuple
        if tuple[1].is_floating_point():
            return tuple[1].dtype

    if last_tuple is not None:
        # fallback to the last dtype
        return last_tuple[1].dtype
149
150


151
class ModelMixin(torch.nn.Module, PushToHubMixin):
152
153
154
    r"""
    Base class for all models.

Steven Liu's avatar
Steven Liu committed
155
156
    [`ModelMixin`] takes care of storing the model configuration and provides methods for loading, downloading and
    saving models.
157

Steven Liu's avatar
Steven Liu committed
158
        - **config_name** ([`str`]) -- Filename to save a model to when calling [`~models.ModelMixin.save_pretrained`].
159
    """
160

161
    config_name = CONFIG_NAME
Patrick von Platen's avatar
Patrick von Platen committed
162
    _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]
163
    _supports_gradient_checkpointing = False
164
    _keys_to_ignore_on_load_unexpected = None
165
    _no_split_modules = None
166
    _keep_in_fp32_modules = None
Aryan's avatar
Aryan committed
167
    _skip_layerwise_casting_patterns = None
168

169
    def __init__(self):
170
171
        super().__init__()

172
173
        self._gradient_checkpointing_func = None

174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
    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)

192
193
194
195
196
197
198
    @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())

199
    def enable_gradient_checkpointing(self, gradient_checkpointing_func: Optional[Callable] = None) -> None:
200
        """
Steven Liu's avatar
Steven Liu committed
201
202
        Activates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
        *checkpoint activations* in other frameworks).
203
204
205
206
207

        Args:
            gradient_checkpointing_func (`Callable`, *optional*):
                The function to use for gradient checkpointing. If `None`, the default PyTorch checkpointing function
                is used (`torch.utils.checkpoint.checkpoint`).
208
209
        """
        if not self._supports_gradient_checkpointing:
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
            raise ValueError(
                f"{self.__class__.__name__} does not support gradient checkpointing. Please make sure to set the boolean attribute "
                f"`_supports_gradient_checkpointing` to `True` in the class definition."
            )

        if gradient_checkpointing_func is None:

            def _gradient_checkpointing_func(module, *args):
                ckpt_kwargs = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                return torch.utils.checkpoint.checkpoint(
                    module.__call__,
                    *args,
                    **ckpt_kwargs,
                )

            gradient_checkpointing_func = _gradient_checkpointing_func

        self._set_gradient_checkpointing(enable=True, gradient_checkpointing_func=gradient_checkpointing_func)
228

229
    def disable_gradient_checkpointing(self) -> None:
230
        """
Steven Liu's avatar
Steven Liu committed
231
232
        Deactivates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
        *checkpoint activations* in other frameworks).
233
234
        """
        if self._supports_gradient_checkpointing:
235
            self._set_gradient_checkpointing(enable=False)
236

237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
    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)

Juan Acevedo's avatar
Juan Acevedo committed
267
    def set_use_xla_flash_attention(
268
        self, use_xla_flash_attention: bool, partition_spec: Optional[Callable] = None, **kwargs
Juan Acevedo's avatar
Juan Acevedo committed
269
270
271
272
273
274
    ) -> None:
        # Recursively walk through all the children.
        # Any children which exposes the set_use_xla_flash_attention method
        # gets the message
        def fn_recursive_set_flash_attention(module: torch.nn.Module):
            if hasattr(module, "set_use_xla_flash_attention"):
275
                module.set_use_xla_flash_attention(use_xla_flash_attention, partition_spec, **kwargs)
Juan Acevedo's avatar
Juan Acevedo committed
276
277
278
279
280
281
282
283

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

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

284
    def enable_xla_flash_attention(self, partition_spec: Optional[Callable] = None, **kwargs):
Juan Acevedo's avatar
Juan Acevedo committed
285
286
287
        r"""
        Enable the flash attention pallals kernel for torch_xla.
        """
288
        self.set_use_xla_flash_attention(True, partition_spec, **kwargs)
Juan Acevedo's avatar
Juan Acevedo committed
289
290
291
292
293
294
295

    def disable_xla_flash_attention(self):
        r"""
        Disable the flash attention pallals kernel for torch_xla.
        """
        self.set_use_xla_flash_attention(False)

296
297
298
    def set_use_memory_efficient_attention_xformers(
        self, valid: bool, attention_op: Optional[Callable] = None
    ) -> None:
299
300
301
302
303
        # 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"):
304
                module.set_use_memory_efficient_attention_xformers(valid, attention_op)
305
306
307
308
309
310
311
312

            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)

313
    def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None) -> None:
314
        r"""
Steven Liu's avatar
Steven Liu committed
315
        Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
316

Steven Liu's avatar
Steven Liu committed
317
318
        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.
319

Steven Liu's avatar
Steven Liu committed
320
321
322
323
324
325
        <Tip warning={true}>

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

        </Tip>
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345

        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)
        ```
346
        """
347
        self.set_use_memory_efficient_attention_xformers(True, attention_op)
348

349
    def disable_xformers_memory_efficient_attention(self) -> None:
350
        r"""
Steven Liu's avatar
Steven Liu committed
351
        Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
352
353
354
        """
        self.set_use_memory_efficient_attention_xformers(False)

Aryan's avatar
Aryan committed
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
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
    def enable_layerwise_casting(
        self,
        storage_dtype: torch.dtype = torch.float8_e4m3fn,
        compute_dtype: Optional[torch.dtype] = None,
        skip_modules_pattern: Optional[Tuple[str, ...]] = None,
        skip_modules_classes: Optional[Tuple[Type[torch.nn.Module], ...]] = None,
        non_blocking: bool = False,
    ) -> None:
        r"""
        Activates layerwise casting for the current model.

        Layerwise casting is a technique that casts the model weights to a lower precision dtype for storage but
        upcasts them on-the-fly to a higher precision dtype for computation. This process can significantly reduce the
        memory footprint from model weights, but may lead to some quality degradation in the outputs. Most degradations
        are negligible, mostly stemming from weight casting in normalization and modulation layers.

        By default, most models in diffusers set the `_skip_layerwise_casting_patterns` attribute to ignore patch
        embedding, positional embedding and normalization layers. This is because these layers are most likely
        precision-critical for quality. If you wish to change this behavior, you can set the
        `_skip_layerwise_casting_patterns` attribute to `None`, or call
        [`~hooks.layerwise_casting.apply_layerwise_casting`] with custom arguments.

        Example:
            Using [`~models.ModelMixin.enable_layerwise_casting`]:

            ```python
            >>> from diffusers import CogVideoXTransformer3DModel

            >>> transformer = CogVideoXTransformer3DModel.from_pretrained(
            ...     "THUDM/CogVideoX-5b", subfolder="transformer", torch_dtype=torch.bfloat16
            ... )

            >>> # Enable layerwise casting via the model, which ignores certain modules by default
            >>> transformer.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)
            ```

        Args:
            storage_dtype (`torch.dtype`):
                The dtype to which the model should be cast for storage.
            compute_dtype (`torch.dtype`):
                The dtype to which the model weights should be cast during the forward pass.
            skip_modules_pattern (`Tuple[str, ...]`, *optional*):
                A list of patterns to match the names of the modules to skip during the layerwise casting process. If
                set to `None`, default skip patterns are used to ignore certain internal layers of modules and PEFT
                layers.
            skip_modules_classes (`Tuple[Type[torch.nn.Module], ...]`, *optional*):
                A list of module classes to skip during the layerwise casting process.
            non_blocking (`bool`, *optional*, defaults to `False`):
                If `True`, the weight casting operations are non-blocking.
        """

        user_provided_patterns = True
        if skip_modules_pattern is None:
            from ..hooks.layerwise_casting import DEFAULT_SKIP_MODULES_PATTERN

            skip_modules_pattern = DEFAULT_SKIP_MODULES_PATTERN
            user_provided_patterns = False
        if self._keep_in_fp32_modules is not None:
            skip_modules_pattern += tuple(self._keep_in_fp32_modules)
        if self._skip_layerwise_casting_patterns is not None:
            skip_modules_pattern += tuple(self._skip_layerwise_casting_patterns)
        skip_modules_pattern = tuple(set(skip_modules_pattern))

        if is_peft_available() and not user_provided_patterns:
            # By default, we want to skip all peft layers because they have a very low memory footprint.
            # If users want to apply layerwise casting on peft layers as well, they can utilize the
            # `~diffusers.hooks.layerwise_casting.apply_layerwise_casting` function which provides
            # them with more flexibility and control.

            from peft.tuners.loha.layer import LoHaLayer
            from peft.tuners.lokr.layer import LoKrLayer
            from peft.tuners.lora.layer import LoraLayer

            for layer in (LoHaLayer, LoKrLayer, LoraLayer):
                skip_modules_pattern += tuple(layer.adapter_layer_names)

        if compute_dtype is None:
            logger.info("`compute_dtype` not provided when enabling layerwise casting. Using dtype of the model.")
            compute_dtype = self.dtype

        apply_layerwise_casting(
            self, storage_dtype, compute_dtype, skip_modules_pattern, skip_modules_classes, non_blocking
        )

439
440
441
442
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
        is_main_process: bool = True,
443
        save_function: Optional[Callable] = None,
444
        safe_serialization: bool = True,
445
        variant: Optional[str] = None,
446
        max_shard_size: Union[int, str] = "10GB",
447
448
        push_to_hub: bool = False,
        **kwargs,
449
450
    ):
        """
Steven Liu's avatar
Steven Liu committed
451
452
        Save a model and its configuration file to a directory so that it can be reloaded using the
        [`~models.ModelMixin.from_pretrained`] class method.
453
454
455

        Arguments:
            save_directory (`str` or `os.PathLike`):
Steven Liu's avatar
Steven Liu committed
456
                Directory to save a model and its configuration file to. Will be created if it doesn't exist.
457
            is_main_process (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
458
459
460
                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.
461
            save_function (`Callable`):
Steven Liu's avatar
Steven Liu committed
462
463
                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
464
                `DIFFUSERS_SAVE_MODE`.
465
            safe_serialization (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
466
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
467
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
468
                If specified, weights are saved in the format `pytorch_model.<variant>.bin`.
469
            max_shard_size (`int` or `str`, defaults to `"10GB"`):
470
471
                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"`).
472
473
474
475
                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).
476
477
478
479
480
481
            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.
482
483
484
485
486
        """
        if os.path.isfile(save_directory):
            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
            return

487
488
489
490
491
492
493
494
495
496
497
498
499
        hf_quantizer = getattr(self, "hf_quantizer", None)
        if hf_quantizer is not None:
            quantization_serializable = (
                hf_quantizer is not None
                and isinstance(hf_quantizer, DiffusersQuantizer)
                and hf_quantizer.is_serializable
            )
            if not quantization_serializable:
                raise ValueError(
                    f"The model is quantized with {hf_quantizer.quantization_config.quant_method} and is not serializable - check out the warnings from"
                    " the logger on the traceback to understand the reason why the quantized model is not serializable."
                )

500
501
        weights_name = SAFETENSORS_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
        weights_name = _add_variant(weights_name, variant)
502
503
504
        weights_name_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(
            ".safetensors", "{suffix}.safetensors"
        )
505

506
507
        os.makedirs(save_directory, exist_ok=True)

508
509
        if push_to_hub:
            commit_message = kwargs.pop("commit_message", None)
510
            private = kwargs.pop("private", None)
511
512
513
514
515
516
            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
517
518
519
520
521
        model_to_save = self

        # Attach architecture to the config
        # Save the config
        if is_main_process:
522
            model_to_save.save_config(save_directory)
523
524
525
526
527

        # Save the model
        state_dict = model_to_save.state_dict()

        # Save the model
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
        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}."
575
576
            )
        else:
577
578
            path_to_weights = os.path.join(save_directory, weights_name)
            logger.info(f"Model weights saved in {path_to_weights}")
579

580
        if push_to_hub:
581
582
583
            # 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)
584
            model_card.save(Path(save_directory, "README.md").as_posix())
585

586
587
588
589
590
591
592
593
            self._upload_folder(
                save_directory,
                repo_id,
                token=token,
                commit_message=commit_message,
                create_pr=create_pr,
            )

594
595
596
597
598
599
600
601
602
603
604
605
    def dequantize(self):
        """
        Potentially dequantize the model in case it has been quantized by a quantization method that support
        dequantization.
        """
        hf_quantizer = getattr(self, "hf_quantizer", None)

        if hf_quantizer is None:
            raise ValueError("You need to first quantize your model in order to dequantize it")

        return hf_quantizer.dequantize(self)

606
    @classmethod
607
    @validate_hf_hub_args
608
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
609
        r"""
Steven Liu's avatar
Steven Liu committed
610
        Instantiate a pretrained PyTorch model from a pretrained model configuration.
611

Steven Liu's avatar
Steven Liu committed
612
613
        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()`.
614
615
616
617
618

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

Steven Liu's avatar
Steven Liu committed
619
620
621
622
                    - 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`].
623
624

            cache_dir (`Union[str, os.PathLike]`, *optional*):
Steven Liu's avatar
Steven Liu committed
625
626
                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
627
            torch_dtype (`str` or `torch.dtype`, *optional*):
Steven Liu's avatar
Steven Liu committed
628
629
                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.
630
631
632
633
            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.
            proxies (`Dict[str, str]`, *optional*):
Steven Liu's avatar
Steven Liu committed
634
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
635
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
Steven Liu's avatar
Steven Liu committed
636
            output_loading_info (`bool`, *optional*, defaults to `False`):
637
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
638
            local_files_only(`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
639
640
                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.
641
            token (`str` or *bool*, *optional*):
Steven Liu's avatar
Steven Liu committed
642
643
                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.
644
            revision (`str`, *optional*, defaults to `"main"`):
Steven Liu's avatar
Steven Liu committed
645
646
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
647
648
            from_flax (`bool`, *optional*, defaults to `False`):
                Load the model weights from a Flax checkpoint save file.
649
            subfolder (`str`, *optional*, defaults to `""`):
Steven Liu's avatar
Steven Liu committed
650
                The subfolder location of a model file within a larger model repository on the Hub or locally.
651
            mirror (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
652
653
654
                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.
655
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
Steven Liu's avatar
Steven Liu committed
656
657
                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
658
                same device. Defaults to `None`, meaning that the model will be loaded on CPU.
659

Steven Liu's avatar
Steven Liu committed
660
                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
661
662
                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).
663
            max_memory (`Dict`, *optional*):
Steven Liu's avatar
Steven Liu committed
664
665
                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.
666
            offload_folder (`str` or `os.PathLike`, *optional*):
Steven Liu's avatar
Steven Liu committed
667
                The path to offload weights if `device_map` contains the value `"disk"`.
668
            offload_state_dict (`bool`, *optional*):
Steven Liu's avatar
Steven Liu committed
669
670
671
                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.
672
            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
Steven Liu's avatar
Steven Liu committed
673
674
675
676
                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.
677
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
678
679
                Load weights from a specified `variant` filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.
680
            use_safetensors (`bool`, *optional*, defaults to `None`):
Steven Liu's avatar
Steven Liu committed
681
682
683
                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.
684
685
686
            disable_mmap ('bool', *optional*, defaults to 'False'):
                Whether to disable mmap when loading a Safetensors model. This option can perform better when the model
                is on a network mount or hard drive, which may not handle the seeky-ness of mmap very well.
687
688
689

        <Tip>

Steven Liu's avatar
Steven Liu committed
690
691
692
693
        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.
694
695
696

        </Tip>

Steven Liu's avatar
Steven Liu committed
697
        Example:
698

Steven Liu's avatar
Steven Liu committed
699
700
        ```py
        from diffusers import UNet2DConditionModel
701

Steven Liu's avatar
Steven Liu committed
702
703
704
705
        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:
706

Steven Liu's avatar
Steven Liu committed
707
708
709
710
711
        ```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.
        ```
712
        """
713
        cache_dir = kwargs.pop("cache_dir", None)
714
715
        ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
        force_download = kwargs.pop("force_download", False)
716
        from_flax = kwargs.pop("from_flax", False)
717
718
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
719
720
        local_files_only = kwargs.pop("local_files_only", None)
        token = kwargs.pop("token", None)
721
        revision = kwargs.pop("revision", None)
722
        torch_dtype = kwargs.pop("torch_dtype", None)
Patrick von Platen's avatar
Patrick von Platen committed
723
        subfolder = kwargs.pop("subfolder", None)
724
        device_map = kwargs.pop("device_map", None)
725
726
727
        max_memory = kwargs.pop("max_memory", None)
        offload_folder = kwargs.pop("offload_folder", None)
        offload_state_dict = kwargs.pop("offload_state_dict", False)
728
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
729
        variant = kwargs.pop("variant", None)
730
        use_safetensors = kwargs.pop("use_safetensors", None)
731
        quantization_config = kwargs.pop("quantization_config", None)
Marc Sun's avatar
Marc Sun committed
732
        dduf_entries: Optional[Dict[str, DDUFEntry]] = kwargs.pop("dduf_entries", None)
733
        disable_mmap = kwargs.pop("disable_mmap", False)
734
735
736

        allow_pickle = False
        if use_safetensors is None:
737
            use_safetensors = True
738
            allow_pickle = True
739

740
741
        if low_cpu_mem_usage and not is_accelerate_available():
            low_cpu_mem_usage = False
742
            logger.warning(
743
744
745
746
747
748
749
750
751
752
753
754
                "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`."
            )

755
756
        # 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"):
757
758
759
760
            raise NotImplementedError(
                "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
                " `device_map=None`."
            )
761

762
763
764
765
766
767
768
769
770
771
772
        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`."
            )
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
        # 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.")

804
805
806
        # Load config if we don't provide a configuration
        config_path = pretrained_model_name_or_path

807
808
809
810
811
        user_agent = {
            "diffusers": __version__,
            "file_type": "model",
            "framework": "pytorch",
        }
812

813
814
815
816
817
818
819
820
821
        # 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,
            proxies=proxies,
            local_files_only=local_files_only,
822
            token=token,
823
824
825
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
Marc Sun's avatar
Marc Sun committed
826
            dduf_entries=dduf_entries,
827
828
            **kwargs,
        )
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
        # no in-place modification of the original config.
        config = copy.deepcopy(config)

        # determine initial quantization config.
        #######################################
        pre_quantized = "quantization_config" in config and config["quantization_config"] is not None
        if pre_quantized or quantization_config is not None:
            if pre_quantized:
                config["quantization_config"] = DiffusersAutoQuantizer.merge_quantization_configs(
                    config["quantization_config"], quantization_config
                )
            else:
                config["quantization_config"] = quantization_config
            hf_quantizer = DiffusersAutoQuantizer.from_config(
                config["quantization_config"], pre_quantized=pre_quantized
            )
        else:
            hf_quantizer = None

        if hf_quantizer is not None:
849
            if device_map is not None:
850
                raise NotImplementedError(
851
                    "Currently, providing `device_map` is not supported for quantized models. Providing `device_map` as an input will be added in the future."
852
                )
Aryan's avatar
Aryan committed
853

854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
            hf_quantizer.validate_environment(torch_dtype=torch_dtype, from_flax=from_flax, device_map=device_map)
            torch_dtype = hf_quantizer.update_torch_dtype(torch_dtype)

            # In order to ensure popular quantization methods are supported. Can be disable with `disable_telemetry`
            user_agent["quant"] = hf_quantizer.quantization_config.quant_method.value

            # Force-set to `True` for more mem efficiency
            if low_cpu_mem_usage is None:
                low_cpu_mem_usage = True
                logger.info("Set `low_cpu_mem_usage` to True as `hf_quantizer` is not None.")
            elif not low_cpu_mem_usage:
                raise ValueError("`low_cpu_mem_usage` cannot be False or None when using quantization.")

        # Check if `_keep_in_fp32_modules` is not None
        use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and (
            (torch_dtype == torch.float16) or hasattr(hf_quantizer, "use_keep_in_fp32_modules")
        )
        if use_keep_in_fp32_modules:
            keep_in_fp32_modules = cls._keep_in_fp32_modules
            if not isinstance(keep_in_fp32_modules, list):
                keep_in_fp32_modules = [keep_in_fp32_modules]

            if low_cpu_mem_usage is None:
                low_cpu_mem_usage = True
                logger.info("Set `low_cpu_mem_usage` to True as `_keep_in_fp32_modules` is not None.")
            elif not low_cpu_mem_usage:
                raise ValueError("`low_cpu_mem_usage` cannot be False when `keep_in_fp32_modules` is True.")
        else:
            keep_in_fp32_modules = []
        #######################################
884

885
886
887
888
        # 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)
889
890
891
892
893
894
895
896
897
898
899
900
901
902
        index_file_kwargs = {
            "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,
            "proxies": proxies,
            "local_files_only": local_files_only,
            "token": token,
            "revision": revision,
            "user_agent": user_agent,
            "commit_hash": commit_hash,
Marc Sun's avatar
Marc Sun committed
903
            "dduf_entries": dduf_entries,
904
905
906
907
908
909
        }
        index_file = _fetch_index_file(**index_file_kwargs)
        # In case the index file was not found we still have to consider the legacy format.
        # this becomes applicable when the variant is not None.
        if variant is not None and (index_file is None or not os.path.exists(index_file)):
            index_file = _fetch_index_file_legacy(**index_file_kwargs)
Marc Sun's avatar
Marc Sun committed
910
        if index_file is not None and (dduf_entries or index_file.is_file()):
911
912
913
914
915
            is_sharded = True

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

916
        # load model
917
        model_file = None
918
        if from_flax:
919
            model_file = _get_model_file(
920
                pretrained_model_name_or_path,
921
                weights_name=FLAX_WEIGHTS_NAME,
922
923
924
925
                cache_dir=cache_dir,
                force_download=force_download,
                proxies=proxies,
                local_files_only=local_files_only,
926
                token=token,
927
928
929
                revision=revision,
                subfolder=subfolder,
                user_agent=user_agent,
930
                commit_hash=commit_hash,
931
932
            )
            model = cls.from_config(config, **unused_kwargs)
933

934
935
936
937
938
            # 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:
Marc Sun's avatar
Marc Sun committed
939
            # in the case it is sharded, we have already the index
940
941
942
943
944
945
946
947
948
949
950
            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,
                    local_files_only=local_files_only,
                    token=token,
                    user_agent=user_agent,
                    revision=revision,
                    subfolder=subfolder or "",
Marc Sun's avatar
Marc Sun committed
951
                    dduf_entries=dduf_entries,
952
                )
953
                # TODO: https://github.com/huggingface/diffusers/issues/10013
Marc Sun's avatar
Marc Sun committed
954
955
956
957
                if hf_quantizer is not None or dduf_entries:
                    model_file = _merge_sharded_checkpoints(
                        sharded_ckpt_cached_folder, sharded_metadata, dduf_entries=dduf_entries
                    )
958
959
                    logger.info("Merged sharded checkpoints as `hf_quantizer` is not None.")
                    is_sharded = False
960
961

            elif use_safetensors and not is_sharded:
962
                try:
963
                    model_file = _get_model_file(
964
                        pretrained_model_name_or_path,
965
                        weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant),
966
967
968
969
                        cache_dir=cache_dir,
                        force_download=force_download,
                        proxies=proxies,
                        local_files_only=local_files_only,
970
                        token=token,
971
972
973
                        revision=revision,
                        subfolder=subfolder,
                        user_agent=user_agent,
974
                        commit_hash=commit_hash,
Marc Sun's avatar
Marc Sun committed
975
                        dduf_entries=dduf_entries,
976
                    )
977

978
                except IOError as e:
979
                    logger.error(f"An error occurred while trying to fetch {pretrained_model_name_or_path}: {e}")
980
                    if not allow_pickle:
981
982
983
984
985
986
                        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:
987
                model_file = _get_model_file(
988
                    pretrained_model_name_or_path,
989
                    weights_name=_add_variant(WEIGHTS_NAME, variant),
990
991
992
993
                    cache_dir=cache_dir,
                    force_download=force_download,
                    proxies=proxies,
                    local_files_only=local_files_only,
994
                    token=token,
995
996
997
                    revision=revision,
                    subfolder=subfolder,
                    user_agent=user_agent,
998
                    commit_hash=commit_hash,
Marc Sun's avatar
Marc Sun committed
999
                    dduf_entries=dduf_entries,
1000
1001
1002
1003
1004
1005
1006
                )

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

1007
1008
1009
1010
1011
                if hf_quantizer is not None:
                    hf_quantizer.preprocess_model(
                        model=model, device_map=device_map, keep_in_fp32_modules=keep_in_fp32_modules
                    )

1012
                # if device_map is None, load the state dict and move the params from meta device to the cpu
1013
                if device_map is None and not is_sharded:
1014
1015
1016
1017
1018
                    # `torch.cuda.current_device()` is fine here when `hf_quantizer` is not None.
                    # It would error out during the `validate_environment()` call above in the absence of cuda.
                    if hf_quantizer is None:
                        param_device = "cpu"
                    # TODO (sayakpaul,  SunMarc): remove this after model loading refactor
Aryan's avatar
Aryan committed
1019
                    else:
1020
                        param_device = torch.device(torch.cuda.current_device())
Marc Sun's avatar
Marc Sun committed
1021
1022
1023
                    state_dict = load_state_dict(
                        model_file, variant=variant, dduf_entries=dduf_entries, disable_mmap=disable_mmap
                    )
1024
                    model._convert_deprecated_attention_blocks(state_dict)
1025

1026
                    # move the params from meta device to cpu
1027
                    missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())
1028
1029
                    if hf_quantizer is not None:
                        missing_keys = hf_quantizer.update_missing_keys(model, missing_keys, prefix="")
1030
1031
1032
1033
                    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
1034
                            " `low_cpu_mem_usage=False` and `device_map=None` if you want to randomly initialize"
1035
1036
                            " those weights or else make sure your checkpoint file is correct."
                        )
1037

hlky's avatar
hlky committed
1038
1039
                    named_buffers = model.named_buffers()

1040
1041
1042
1043
1044
1045
                    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,
1046
1047
                        hf_quantizer=hf_quantizer,
                        keep_in_fp32_modules=keep_in_fp32_modules,
hlky's avatar
hlky committed
1048
                        named_buffers=named_buffers,
1049
                    )
1050
1051
1052
1053
1054
1055

                    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:
1056
                        logger.warning(
1057
1058
1059
                            f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
                        )

1060
1061
                else:  # else let accelerate handle loading and dispatching.
                    # Load weights and dispatch according to the device_map
Alexander Pivovarov's avatar
Alexander Pivovarov committed
1062
                    # by default the device_map is None and the weights are loaded on the CPU
1063
1064
1065
                    device_map = _determine_device_map(
                        model, device_map, max_memory, torch_dtype, keep_in_fp32_modules, hf_quantizer
                    )
1066
1067
1068
                    if device_map is None and is_sharded:
                        # we load the parameters on the cpu
                        device_map = {"": "cpu"}
1069
1070
1071
                    try:
                        accelerate.load_checkpoint_and_dispatch(
                            model,
1072
                            model_file if not is_sharded else index_file,
1073
1074
1075
1076
1077
                            device_map,
                            max_memory=max_memory,
                            offload_folder=offload_folder,
                            offload_state_dict=offload_state_dict,
                            dtype=torch_dtype,
1078
                            strict=True,
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
                        )
                    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):
1091
                            logger.warning(
1092
1093
1094
1095
1096
1097
1098
1099
1100
                                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,
1101
                                model_file if not is_sharded else index_file,
1102
1103
1104
1105
1106
                                device_map,
                                max_memory=max_memory,
                                offload_folder=offload_folder,
                                offload_state_dict=offload_state_dict,
                                dtype=torch_dtype,
1107
                                strict=True,
1108
1109
1110
1111
                            )
                            model._undo_temp_convert_self_to_deprecated_attention_blocks()
                        else:
                            raise e
1112
1113
1114
1115
1116
1117
1118
1119

                loading_info = {
                    "missing_keys": [],
                    "unexpected_keys": [],
                    "mismatched_keys": [],
                    "error_msgs": [],
                }
            else:
1120
                model = cls.from_config(config, **unused_kwargs)
1121

Marc Sun's avatar
Marc Sun committed
1122
1123
1124
                state_dict = load_state_dict(
                    model_file, variant=variant, dduf_entries=dduf_entries, disable_mmap=disable_mmap
                )
1125
                model._convert_deprecated_attention_blocks(state_dict)
1126

1127
1128
1129
1130
1131
1132
1133
                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,
                )
1134

1135
1136
1137
1138
1139
1140
                loading_info = {
                    "missing_keys": missing_keys,
                    "unexpected_keys": unexpected_keys,
                    "mismatched_keys": mismatched_keys,
                    "error_msgs": error_msgs,
                }
1141

1142
1143
1144
1145
        if hf_quantizer is not None:
            hf_quantizer.postprocess_model(model)
            model.hf_quantizer = hf_quantizer

1146
1147
1148
1149
        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)}."
            )
1150
1151
1152
        # When using `use_keep_in_fp32_modules` if we do a global `to()` here, then we will
        # completely lose the effectivity of `use_keep_in_fp32_modules`.
        elif torch_dtype is not None and hf_quantizer is None and not use_keep_in_fp32_modules:
1153
1154
            model = model.to(torch_dtype)

1155
1156
1157
1158
1159
1160
        if hf_quantizer is not None:
            # We also make sure to purge `_pre_quantization_dtype` when we serialize
            # the model config because `_pre_quantization_dtype` is `torch.dtype`, not JSON serializable.
            model.register_to_config(_name_or_path=pretrained_model_name_or_path, _pre_quantization_dtype=torch_dtype)
        else:
            model.register_to_config(_name_or_path=pretrained_model_name_or_path)
1161
1162
1163
1164

        # Set model in evaluation mode to deactivate DropOut modules by default
        model.eval()
        if output_loading_info:
1165
1166
1167
1168
            return model, loading_info

        return model

1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
    # Adapted from `transformers`.
    @wraps(torch.nn.Module.cuda)
    def cuda(self, *args, **kwargs):
        # Checks if the model has been loaded in 4-bit or 8-bit with BNB
        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
            if getattr(self, "is_loaded_in_8bit", False):
                raise ValueError(
                    "Calling `cuda()` is not supported for `8-bit` quantized models. "
                    " Please use the model as it is, since the model has already been set to the correct devices."
                )
            elif is_bitsandbytes_version("<", "0.43.2"):
                raise ValueError(
                    "Calling `cuda()` is not supported for `4-bit` quantized models with the installed version of bitsandbytes. "
                    f"The current device is `{self.device}`. If you intended to move the model, please install bitsandbytes >= 0.43.2."
                )
        return super().cuda(*args, **kwargs)

    # Adapted from `transformers`.
    @wraps(torch.nn.Module.to)
    def to(self, *args, **kwargs):
        dtype_present_in_args = "dtype" in kwargs

        if not dtype_present_in_args:
            for arg in args:
                if isinstance(arg, torch.dtype):
                    dtype_present_in_args = True
                    break

1197
        if getattr(self, "is_quantized", False):
1198
1199
            if dtype_present_in_args:
                raise ValueError(
1200
1201
                    "Casting a quantized model to a new `dtype` is unsupported. To set the dtype of unquantized layers, please "
                    "use the `torch_dtype` argument when loading the model using `from_pretrained` or `from_single_file`"
1202
1203
                )

1204
        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
            if getattr(self, "is_loaded_in_8bit", False):
                raise ValueError(
                    "`.to` is not supported for `8-bit` bitsandbytes models. Please use the model as it is, since the"
                    " model has already been set to the correct devices and casted to the correct `dtype`."
                )
            elif is_bitsandbytes_version("<", "0.43.2"):
                raise ValueError(
                    "Calling `to()` is not supported for `4-bit` quantized models with the installed version of bitsandbytes. "
                    f"The current device is `{self.device}`. If you intended to move the model, please install bitsandbytes >= 0.43.2."
                )
        return super().to(*args, **kwargs)

    # Taken from `transformers`.
    def half(self, *args):
        # Checks if the model is quantized
        if getattr(self, "is_quantized", False):
            raise ValueError(
                "`.half()` is not supported for quantized model. Please use the model as it is, since the"
                " model has already been cast to the correct `dtype`."
            )
        else:
            return super().half(*args)

    # Taken from `transformers`.
    def float(self, *args):
        # Checks if the model is quantized
        if getattr(self, "is_quantized", False):
            raise ValueError(
                "`.float()` is not supported for quantized model. Please use the model as it is, since the"
                " model has already been cast to the correct `dtype`."
            )
        else:
            return super().float(*args)

1239
1240
1241
1242
    @classmethod
    def _load_pretrained_model(
        cls,
        model,
1243
        state_dict: OrderedDict,
1244
        resolved_archive_file,
1245
1246
        pretrained_model_name_or_path: Union[str, os.PathLike],
        ignore_mismatched_sizes: bool = False,
1247
1248
1249
    ):
        # Retrieve missing & unexpected_keys
        model_state_dict = model.state_dict()
1250
        loaded_keys = list(state_dict.keys())
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300

        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
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
        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."
            )
1340
1341
1342

        return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs

1343
1344
1345
1346
1347
1348
1349
1350
1351
    @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

1352
1353
1354
    # Adapted from `transformers` modeling_utils.py
    def _get_no_split_modules(self, device_map: str):
        """
1355
        Get the modules of the model that should not be split when using device_map. We iterate through the modules to
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
        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)

1382
    @property
1383
    def device(self) -> torch.device:
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
        """
        `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
1399
        Get number of (trainable or non-embedding) parameters in the module.
1400
1401
1402

        Args:
            only_trainable (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
1403
                Whether or not to return only the number of trainable parameters.
1404
            exclude_embeddings (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
1405
                Whether or not to return only the number of non-embedding parameters.
1406
1407
1408

        Returns:
            `int`: The number of parameters.
Steven Liu's avatar
Steven Liu committed
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419

        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
        ```
1420
        """
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
        is_loaded_in_4bit = getattr(self, "is_loaded_in_4bit", False)

        if is_loaded_in_4bit:
            if is_bitsandbytes_available():
                import bitsandbytes as bnb
            else:
                raise ValueError(
                    "bitsandbytes is not installed but it seems that the model has been loaded in 4bit precision, something went wrong"
                    " make sure to install bitsandbytes with `pip install bitsandbytes`. You also need a GPU. "
                )
1431
1432
1433

        if exclude_embeddings:
            embedding_param_names = [
1434
                f"{name}.weight" for name, module_type in self.named_modules() if isinstance(module_type, nn.Embedding)
1435
            ]
1436
            total_parameters = [
1437
1438
1439
                parameter for name, parameter in self.named_parameters() if name not in embedding_param_names
            ]
        else:
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
            total_parameters = list(self.parameters())

        total_numel = []

        for param in total_parameters:
            if param.requires_grad or not only_trainable:
                # For 4bit models, we need to multiply the number of parameters by 2 as half of the parameters are
                # used for the 4bit quantization (uint8 tensors are stored)
                if is_loaded_in_4bit and isinstance(param, bnb.nn.Params4bit):
                    if hasattr(param, "element_size"):
                        num_bytes = param.element_size()
                    elif hasattr(param, "quant_storage"):
                        num_bytes = param.quant_storage.itemsize
                    else:
                        num_bytes = 1
                    total_numel.append(param.numel() * 2 * num_bytes)
                else:
                    total_numel.append(param.numel())

        return sum(total_numel)

    def get_memory_footprint(self, return_buffers=True):
        r"""
        Get the memory footprint of a model. This will return the memory footprint of the current model in bytes.
        Useful to benchmark the memory footprint of the current model and design some tests. Solution inspired from the
        PyTorch discussions: https://discuss.pytorch.org/t/gpu-memory-that-model-uses/56822/2

        Arguments:
            return_buffers (`bool`, *optional*, defaults to `True`):
                Whether to return the size of the buffer tensors in the computation of the memory footprint. Buffers
                are tensors that do not require gradients and not registered as parameters. E.g. mean and std in batch
                norm layers. Please see: https://discuss.pytorch.org/t/what-pytorch-means-by-buffers/120266/2
        """
        mem = sum([param.nelement() * param.element_size() for param in self.parameters()])
        if return_buffers:
            mem_bufs = sum([buf.nelement() * buf.element_size() for buf in self.buffers()])
            mem = mem + mem_bufs
        return mem
1478

1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
    def _set_gradient_checkpointing(
        self, enable: bool = True, gradient_checkpointing_func: Callable = torch.utils.checkpoint.checkpoint
    ) -> None:
        is_gradient_checkpointing_set = False

        for name, module in self.named_modules():
            if hasattr(module, "gradient_checkpointing"):
                logger.debug(f"Setting `gradient_checkpointing={enable}` for '{name}'")
                module._gradient_checkpointing_func = gradient_checkpointing_func
                module.gradient_checkpointing = enable
                is_gradient_checkpointing_set = True

        if not is_gradient_checkpointing_set:
            raise ValueError(
                f"The module {self.__class__.__name__} does not support gradient checkpointing. Please make sure to "
                f"use a module that supports gradient checkpointing by creating a boolean attribute `gradient_checkpointing`."
            )

1497
    def _convert_deprecated_attention_blocks(self, state_dict: OrderedDict) -> None:
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
        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")
1540

1541
    def _temp_convert_self_to_deprecated_attention_blocks(self) -> None:
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
        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

1568
    def _undo_temp_convert_self_to_deprecated_attention_blocks(self) -> None:
1569
1570
        deprecated_attention_block_modules = []

1571
        def recursive_find_attn_block(module) -> None:
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
            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
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600


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):
1601
        # To prevent dependency import problem.
1602
1603
        from .model_loading_utils import _fetch_remapped_cls_from_config

1604
1605
1606
        # Create a copy of the kwargs so that we don't mess with the keyword arguments in the downstream calls.
        kwargs_copy = kwargs.copy()

1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
        cache_dir = kwargs.pop("cache_dir", None)
        force_download = kwargs.pop("force_download", False)
        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,
            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)

1642
        return remapped_class.from_pretrained(pretrained_model_name_or_path, **kwargs_copy)