modeling_utils.py 87.4 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
24
import shutil
import tempfile
25
from collections import OrderedDict
26
from contextlib import ExitStack, contextmanager
27
from functools import wraps
28
from pathlib import Path
29
from typing import Any, Callable, ContextManager, Dict, List, Optional, Tuple, Type, Union
30

31
import safetensors
32
import torch
33
import torch.utils.checkpoint
Marc Sun's avatar
Marc Sun committed
34
from huggingface_hub import DDUFEntry, create_repo, split_torch_state_dict_into_shards
35
from huggingface_hub.utils import validate_hf_hub_args
36
from torch import Tensor, nn
37
from typing_extensions import Self
38

39
from .. import __version__
40
41
from ..quantizers import DiffusersAutoQuantizer, DiffusersQuantizer
from ..quantizers.quantization_config import QuantizationMethod
42
from ..utils import (
43
    CONFIG_NAME,
44
    FLAX_WEIGHTS_NAME,
45
    SAFE_WEIGHTS_INDEX_NAME,
46
    SAFETENSORS_WEIGHTS_NAME,
47
    WEIGHTS_INDEX_NAME,
48
    WEIGHTS_NAME,
49
    _add_variant,
50
    _get_checkpoint_shard_files,
51
    _get_model_file,
52
    deprecate,
53
    is_accelerate_available,
54
55
    is_bitsandbytes_available,
    is_bitsandbytes_version,
Aryan's avatar
Aryan committed
56
    is_peft_available,
57
58
59
    is_torch_version,
    logging,
)
60
61
62
63
64
from ..utils.hub_utils import (
    PushToHubMixin,
    load_or_create_model_card,
    populate_model_card,
)
65
from ..utils.torch_utils import empty_device_cache
66
from .model_loading_utils import (
67
    _caching_allocator_warmup,
68
    _determine_device_map,
69
    _expand_device_map,
70
    _fetch_index_file,
71
    _fetch_index_file_legacy,
72
    _find_mismatched_keys,
73
74
75
76
    _load_state_dict_into_model,
    load_model_dict_into_meta,
    load_state_dict,
)
77
78


79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
class ContextManagers:
    """
    Wrapper for `contextlib.ExitStack` which enters a collection of context managers. Adaptation of `ContextManagers`
    in the `fastcore` library.
    """

    def __init__(self, context_managers: List[ContextManager]):
        self.context_managers = context_managers
        self.stack = ExitStack()

    def __enter__(self):
        for context_manager in self.context_managers:
            self.stack.enter_context(context_manager)

    def __exit__(self, *args, **kwargs):
        self.stack.__exit__(*args, **kwargs)


97
98
logger = logging.get_logger(__name__)

99
100
_REGEX_SHARD = re.compile(r"(.*?)-\d{5}-of-\d{5}")

101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
TORCH_INIT_FUNCTIONS = {
    "uniform_": nn.init.uniform_,
    "normal_": nn.init.normal_,
    "trunc_normal_": nn.init.trunc_normal_,
    "constant_": nn.init.constant_,
    "xavier_uniform_": nn.init.xavier_uniform_,
    "xavier_normal_": nn.init.xavier_normal_,
    "kaiming_uniform_": nn.init.kaiming_uniform_,
    "kaiming_normal_": nn.init.kaiming_normal_,
    "uniform": nn.init.uniform,
    "normal": nn.init.normal,
    "xavier_uniform": nn.init.xavier_uniform,
    "xavier_normal": nn.init.xavier_normal,
    "kaiming_uniform": nn.init.kaiming_uniform,
    "kaiming_normal": nn.init.kaiming_normal,
}
117

118
119
120
121
122
123
if is_torch_version(">=", "1.9.0"):
    _LOW_CPU_MEM_USAGE_DEFAULT = True
else:
    _LOW_CPU_MEM_USAGE_DEFAULT = False


124
125
if is_accelerate_available():
    import accelerate
126
127
    from accelerate import dispatch_model
    from accelerate.utils import load_offloaded_weights, save_offload_index
128
129


130
def get_parameter_device(parameter: torch.nn.Module) -> torch.device:
Aryan's avatar
Aryan committed
131
132
133
134
135
136
137
138
    from ..hooks.group_offloading import _get_group_onload_device

    try:
        # Try to get the onload device from the group offloading hook
        return _get_group_onload_device(parameter)
    except ValueError:
        pass

139
    try:
Aryan's avatar
Aryan committed
140
141
        # If the onload device is not available due to no group offloading hooks, try to get the device
        # from the first parameter or buffer
Patrick von Platen's avatar
Patrick von Platen committed
142
143
        parameters_and_buffers = itertools.chain(parameter.parameters(), parameter.buffers())
        return next(parameters_and_buffers).device
144
145
146
147
148
149
150
151
152
153
154
155
    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


156
def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype:
157
158
159
    """
    Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.
    """
Aryan's avatar
Aryan committed
160
161
162
163
164
165
166
167
168
169
170
    # 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
171
    last_dtype = None
172
173

    for name, param in parameter.named_parameters():
174
        last_dtype = param.dtype
175
176
177
178
179
        if (
            hasattr(parameter, "_keep_in_fp32_modules")
            and parameter._keep_in_fp32_modules
            and any(m in name for m in parameter._keep_in_fp32_modules)
        ):
180
181
            continue

182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
        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
209
210


211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
def check_support_param_buffer_assignment(model_to_load, state_dict, start_prefix=""):
    """
    Checks if `model_to_load` supports param buffer assignment (such as when loading in empty weights) by first
    checking if the model explicitly disables it, then by ensuring that the state dict keys are a subset of the model's
    parameters.

    """
    if model_to_load.device.type == "meta":
        return False

    if len([key for key in state_dict if key.startswith(start_prefix)]) == 0:
        return False

    # Some models explicitly do not support param buffer assignment
    if not getattr(model_to_load, "_supports_param_buffer_assignment", True):
        logger.debug(
            f"{model_to_load.__class__.__name__} does not support param buffer assignment, loading will be slower"
        )
        return False

    # If the model does, the incoming `state_dict` and the `model_to_load` must be the same dtype
    first_key = next(iter(model_to_load.state_dict().keys()))
    if start_prefix + first_key in state_dict:
        return state_dict[start_prefix + first_key].dtype == model_to_load.state_dict()[first_key].dtype

    return False


@contextmanager
def no_init_weights():
    """
    Context manager to globally disable weight initialization to speed up loading large models. To do that, all the
    torch.nn.init function are all replaced with skip.
    """

    def _skip_init(*args, **kwargs):
        pass

    for name, init_func in TORCH_INIT_FUNCTIONS.items():
        setattr(torch.nn.init, name, _skip_init)
    try:
        yield
    finally:
        # Restore the original initialization functions
        for name, init_func in TORCH_INIT_FUNCTIONS.items():
            setattr(torch.nn.init, name, init_func)


259
class ModelMixin(torch.nn.Module, PushToHubMixin):
260
261
262
    r"""
    Base class for all models.

Steven Liu's avatar
Steven Liu committed
263
264
    [`ModelMixin`] takes care of storing the model configuration and provides methods for loading, downloading and
    saving models.
265

Steven Liu's avatar
Steven Liu committed
266
        - **config_name** ([`str`]) -- Filename to save a model to when calling [`~models.ModelMixin.save_pretrained`].
267
    """
268

269
    config_name = CONFIG_NAME
Patrick von Platen's avatar
Patrick von Platen committed
270
    _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]
271
    _supports_gradient_checkpointing = False
272
    _keys_to_ignore_on_load_unexpected = None
273
    _no_split_modules = None
274
    _keep_in_fp32_modules = None
Aryan's avatar
Aryan committed
275
    _skip_layerwise_casting_patterns = None
Aryan's avatar
Aryan committed
276
    _supports_group_offloading = True
277
    _repeated_blocks = []
278

279
    def __init__(self):
280
281
        super().__init__()

282
283
        self._gradient_checkpointing_func = None

284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
    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)

302
303
304
305
306
307
308
    @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())

309
    def enable_gradient_checkpointing(self, gradient_checkpointing_func: Optional[Callable] = None) -> None:
310
        """
Steven Liu's avatar
Steven Liu committed
311
312
        Activates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
        *checkpoint activations* in other frameworks).
313
314
315
316
317

        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`).
318
319
        """
        if not self._supports_gradient_checkpointing:
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
            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)
338

339
    def disable_gradient_checkpointing(self) -> None:
340
        """
Steven Liu's avatar
Steven Liu committed
341
342
        Deactivates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
        *checkpoint activations* in other frameworks).
343
344
        """
        if self._supports_gradient_checkpointing:
345
            self._set_gradient_checkpointing(enable=False)
346

347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
    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
377
    def set_use_xla_flash_attention(
378
        self, use_xla_flash_attention: bool, partition_spec: Optional[Callable] = None, **kwargs
Juan Acevedo's avatar
Juan Acevedo committed
379
380
381
382
383
384
    ) -> 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"):
385
                module.set_use_xla_flash_attention(use_xla_flash_attention, partition_spec, **kwargs)
Juan Acevedo's avatar
Juan Acevedo committed
386
387
388
389
390
391
392
393

            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)

394
    def enable_xla_flash_attention(self, partition_spec: Optional[Callable] = None, **kwargs):
Juan Acevedo's avatar
Juan Acevedo committed
395
396
397
        r"""
        Enable the flash attention pallals kernel for torch_xla.
        """
398
        self.set_use_xla_flash_attention(True, partition_spec, **kwargs)
Juan Acevedo's avatar
Juan Acevedo committed
399
400
401
402
403
404
405

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

406
407
408
    def set_use_memory_efficient_attention_xformers(
        self, valid: bool, attention_op: Optional[Callable] = None
    ) -> None:
409
410
411
412
413
        # 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"):
414
                module.set_use_memory_efficient_attention_xformers(valid, attention_op)
415
416
417
418
419
420
421
422

            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)

423
    def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None) -> None:
424
        r"""
Steven Liu's avatar
Steven Liu committed
425
        Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
426

Steven Liu's avatar
Steven Liu committed
427
428
        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.
429

Steven Liu's avatar
Steven Liu committed
430
431
432
433
434
435
        <Tip warning={true}>

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

        </Tip>
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455

        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)
        ```
456
        """
457
        self.set_use_memory_efficient_attention_xformers(True, attention_op)
458

459
    def disable_xformers_memory_efficient_attention(self) -> None:
460
        r"""
Steven Liu's avatar
Steven Liu committed
461
        Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
462
463
464
        """
        self.set_use_memory_efficient_attention_xformers(False)

Aryan's avatar
Aryan committed
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
    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.
        """
Aryan's avatar
Aryan committed
515
        from ..hooks import apply_layerwise_casting
Aryan's avatar
Aryan committed
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
543
544
545
546
547
548
549

        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
        )

Aryan's avatar
Aryan committed
550
551
552
553
554
555
556
557
    def enable_group_offload(
        self,
        onload_device: torch.device,
        offload_device: torch.device = torch.device("cpu"),
        offload_type: str = "block_level",
        num_blocks_per_group: Optional[int] = None,
        non_blocking: bool = False,
        use_stream: bool = False,
558
        record_stream: bool = False,
559
        low_cpu_mem_usage=False,
560
        offload_to_disk_path: Optional[str] = None,
Aryan's avatar
Aryan committed
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
    ) -> None:
        r"""
        Activates group offloading for the current model.

        See [`~hooks.group_offloading.apply_group_offloading`] for more information.

        Example:

            ```python
            >>> from diffusers import CogVideoXTransformer3DModel

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

            >>> transformer.enable_group_offload(
            ...     onload_device=torch.device("cuda"),
            ...     offload_device=torch.device("cpu"),
            ...     offload_type="leaf_level",
            ...     use_stream=True,
            ... )
            ```
        """
Aryan's avatar
Aryan committed
584
585
        from ..hooks import apply_group_offloading

Aryan's avatar
Aryan committed
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
        if getattr(self, "enable_tiling", None) is not None and getattr(self, "use_tiling", False) and use_stream:
            msg = (
                "Applying group offloading on autoencoders, with CUDA streams, may not work as expected if the first "
                "forward pass is executed with tiling enabled. Please make sure to either:\n"
                "1. Run a forward pass with small input shapes.\n"
                "2. Or, run a forward pass with tiling disabled (can still use small dummy inputs)."
            )
            logger.warning(msg)
        if not self._supports_group_offloading:
            raise ValueError(
                f"{self.__class__.__name__} does not support group offloading. Please make sure to set the boolean attribute "
                f"`_supports_group_offloading` to `True` in the class definition. If you believe this is a mistake, please "
                f"open an issue at https://github.com/huggingface/diffusers/issues."
            )
        apply_group_offloading(
601
602
603
604
605
606
607
608
            module=self,
            onload_device=onload_device,
            offload_device=offload_device,
            offload_type=offload_type,
            num_blocks_per_group=num_blocks_per_group,
            non_blocking=non_blocking,
            use_stream=use_stream,
            record_stream=record_stream,
609
            low_cpu_mem_usage=low_cpu_mem_usage,
610
            offload_to_disk_path=offload_to_disk_path,
Aryan's avatar
Aryan committed
611
612
        )

613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
    def set_attention_backend(self, backend: str) -> None:
        """
        Set the attention backend for the model.

        Args:
            backend (`str`):
                The name of the backend to set. Must be one of the available backends defined in
                `AttentionBackendName`. Available backends can be found in
                `diffusers.attention_dispatch.AttentionBackendName`. Defaults to torch native scaled dot product
                attention as backend.
        """
        from .attention import AttentionModuleMixin
        from .attention_dispatch import AttentionBackendName

        # TODO: the following will not be required when everything is refactored to AttentionModuleMixin
        from .attention_processor import Attention, MochiAttention

        backend = backend.lower()
        available_backends = {x.value for x in AttentionBackendName.__members__.values()}
        if backend not in available_backends:
            raise ValueError(f"`{backend=}` must be one of the following: " + ", ".join(available_backends))

        backend = AttentionBackendName(backend)
        attention_classes = (Attention, MochiAttention, AttentionModuleMixin)

        for module in self.modules():
            if not isinstance(module, attention_classes):
                continue
            processor = module.processor
            if processor is None or not hasattr(processor, "_attention_backend"):
                continue
            processor._attention_backend = backend

    def reset_attention_backend(self) -> None:
        """
        Resets the attention backend for the model. Following calls to `forward` will use the environment default or
        the torch native scaled dot product attention.
        """
        from .attention import AttentionModuleMixin
        from .attention_processor import Attention, MochiAttention

        attention_classes = (Attention, MochiAttention, AttentionModuleMixin)
        for module in self.modules():
            if not isinstance(module, attention_classes):
                continue
            processor = module.processor
            if processor is None or not hasattr(processor, "_attention_backend"):
                continue
            processor._attention_backend = None

663
664
665
666
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
        is_main_process: bool = True,
667
        save_function: Optional[Callable] = None,
668
        safe_serialization: bool = True,
669
        variant: Optional[str] = None,
670
        max_shard_size: Union[int, str] = "10GB",
671
672
        push_to_hub: bool = False,
        **kwargs,
673
674
    ):
        """
Steven Liu's avatar
Steven Liu committed
675
676
        Save a model and its configuration file to a directory so that it can be reloaded using the
        [`~models.ModelMixin.from_pretrained`] class method.
677
678
679

        Arguments:
            save_directory (`str` or `os.PathLike`):
Steven Liu's avatar
Steven Liu committed
680
                Directory to save a model and its configuration file to. Will be created if it doesn't exist.
681
            is_main_process (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
682
683
684
                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.
685
            save_function (`Callable`):
Steven Liu's avatar
Steven Liu committed
686
687
                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
688
                `DIFFUSERS_SAVE_MODE`.
689
            safe_serialization (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
690
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
691
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
692
                If specified, weights are saved in the format `pytorch_model.<variant>.bin`.
693
            max_shard_size (`int` or `str`, defaults to `"10GB"`):
694
695
                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"`).
696
697
698
699
                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).
700
701
702
703
704
705
            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.
706
707
708
709
710
        """
        if os.path.isfile(save_directory):
            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
            return

711
712
713
714
715
716
717
718
719
720
721
722
723
        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."
                )

724
725
        weights_name = SAFETENSORS_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
        weights_name = _add_variant(weights_name, variant)
726
727
728
        weights_name_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(
            ".safetensors", "{suffix}.safetensors"
        )
729

730
731
        os.makedirs(save_directory, exist_ok=True)

732
733
        if push_to_hub:
            commit_message = kwargs.pop("commit_message", None)
734
            private = kwargs.pop("private", None)
735
736
737
738
739
740
            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
741
742
743
744
745
        model_to_save = self

        # Attach architecture to the config
        # Save the config
        if is_main_process:
746
            model_to_save.save_config(save_directory)
747
748
749
750
751

        # Save the model
        state_dict = model_to_save.state_dict()

        # Save the model
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
        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():
775
            shard = {tensor: state_dict[tensor].contiguous() for tensor in tensors}
776
777
778
779
            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.
780
                safetensors.torch.save_file(shard, filepath, metadata={"format": "pt"})
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
            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}."
799
800
            )
        else:
801
802
            path_to_weights = os.path.join(save_directory, weights_name)
            logger.info(f"Model weights saved in {path_to_weights}")
803

804
        if push_to_hub:
805
806
807
            # 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)
808
            model_card.save(Path(save_directory, "README.md").as_posix())
809

810
811
812
813
814
815
816
817
            self._upload_folder(
                save_directory,
                repo_id,
                token=token,
                commit_message=commit_message,
                create_pr=create_pr,
            )

818
819
820
821
822
823
824
825
826
827
828
829
    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)

830
    @classmethod
831
    @validate_hf_hub_args
832
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs) -> Self:
833
        r"""
Steven Liu's avatar
Steven Liu committed
834
        Instantiate a pretrained PyTorch model from a pretrained model configuration.
835

Steven Liu's avatar
Steven Liu committed
836
837
        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()`.
838
839
840
841
842

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

Steven Liu's avatar
Steven Liu committed
843
844
845
846
                    - 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`].
847
848

            cache_dir (`Union[str, os.PathLike]`, *optional*):
Steven Liu's avatar
Steven Liu committed
849
850
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
851
852
            torch_dtype (`torch.dtype`, *optional*):
                Override the default `torch.dtype` and load the model with another dtype.
853
854
855
856
            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
857
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
858
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
Steven Liu's avatar
Steven Liu committed
859
            output_loading_info (`bool`, *optional*, defaults to `False`):
860
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
861
            local_files_only(`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
862
863
                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.
864
            token (`str` or *bool*, *optional*):
Steven Liu's avatar
Steven Liu committed
865
866
                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.
867
            revision (`str`, *optional*, defaults to `"main"`):
Steven Liu's avatar
Steven Liu committed
868
869
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
870
871
            from_flax (`bool`, *optional*, defaults to `False`):
                Load the model weights from a Flax checkpoint save file.
872
            subfolder (`str`, *optional*, defaults to `""`):
Steven Liu's avatar
Steven Liu committed
873
                The subfolder location of a model file within a larger model repository on the Hub or locally.
874
            mirror (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
875
876
877
                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.
878
            device_map (`Union[int, str, torch.device]` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
Steven Liu's avatar
Steven Liu committed
879
880
                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
881
                same device. Defaults to `None`, meaning that the model will be loaded on CPU.
882

883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
                Examples:

                ```py
                >>> from diffusers import AutoModel
                >>> import torch

                >>> # This works.
                >>> model = AutoModel.from_pretrained(
                ...     "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", device_map="cuda"
                ... )
                >>> # This also works (integer accelerator device ID).
                >>> model = AutoModel.from_pretrained(
                ...     "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", device_map=0
                ... )
                >>> # Specifying a supported offloading strategy like "auto" also works.
                >>> model = AutoModel.from_pretrained(
                ...     "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", device_map="auto"
                ... )
                >>> # Specifying a dictionary as `device_map` also works.
                >>> model = AutoModel.from_pretrained(
                ...     "stabilityai/stable-diffusion-xl-base-1.0",
                ...     subfolder="unet",
                ...     device_map={"": torch.device("cuda")},
                ... )
                ```

Steven Liu's avatar
Steven Liu committed
909
                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
910
                more information about each option see [designing a device
911
912
913
914
                map](https://huggingface.co/docs/accelerate/en/concept_guides/big_model_inference#the-devicemap). You
                can also refer to the [Diffusers-specific
                documentation](https://huggingface.co/docs/diffusers/main/en/training/distributed_inference#model-sharding)
                for more concrete examples.
915
            max_memory (`Dict`, *optional*):
Steven Liu's avatar
Steven Liu committed
916
917
                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.
918
            offload_folder (`str` or `os.PathLike`, *optional*):
Steven Liu's avatar
Steven Liu committed
919
                The path to offload weights if `device_map` contains the value `"disk"`.
920
            offload_state_dict (`bool`, *optional*):
Steven Liu's avatar
Steven Liu committed
921
922
923
                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.
924
            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
Steven Liu's avatar
Steven Liu committed
925
926
927
928
                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.
929
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
930
931
                Load weights from a specified `variant` filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.
932
            use_safetensors (`bool`, *optional*, defaults to `None`):
Steven Liu's avatar
Steven Liu committed
933
934
935
                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.
936
937
938
            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.
939
940
941

        <Tip>

Steven Liu's avatar
Steven Liu committed
942
943
944
945
        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.
946
947
948

        </Tip>

Steven Liu's avatar
Steven Liu committed
949
        Example:
950

Steven Liu's avatar
Steven Liu committed
951
952
        ```py
        from diffusers import UNet2DConditionModel
953

Steven Liu's avatar
Steven Liu committed
954
955
956
957
        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:
958

Steven Liu's avatar
Steven Liu committed
959
960
961
962
963
        ```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.
        ```
964
        """
965
        cache_dir = kwargs.pop("cache_dir", None)
966
967
        ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
        force_download = kwargs.pop("force_download", False)
968
        from_flax = kwargs.pop("from_flax", False)
969
970
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
971
972
        local_files_only = kwargs.pop("local_files_only", None)
        token = kwargs.pop("token", None)
973
        revision = kwargs.pop("revision", None)
974
        torch_dtype = kwargs.pop("torch_dtype", None)
Patrick von Platen's avatar
Patrick von Platen committed
975
        subfolder = kwargs.pop("subfolder", None)
976
        device_map = kwargs.pop("device_map", None)
977
978
        max_memory = kwargs.pop("max_memory", None)
        offload_folder = kwargs.pop("offload_folder", None)
979
        offload_state_dict = kwargs.pop("offload_state_dict", None)
980
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
981
        variant = kwargs.pop("variant", None)
982
        use_safetensors = kwargs.pop("use_safetensors", None)
983
        quantization_config = kwargs.pop("quantization_config", None)
Marc Sun's avatar
Marc Sun committed
984
        dduf_entries: Optional[Dict[str, DDUFEntry]] = kwargs.pop("dduf_entries", None)
985
        disable_mmap = kwargs.pop("disable_mmap", False)
986

987
        if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
988
989
990
991
992
            torch_dtype = torch.float32
            logger.warning(
                f"Passed `torch_dtype` {torch_dtype} is not a `torch.dtype`. Defaulting to `torch.float32`."
            )

993
994
        allow_pickle = False
        if use_safetensors is None:
995
            use_safetensors = True
996
            allow_pickle = True
997

998
999
        if low_cpu_mem_usage and not is_accelerate_available():
            low_cpu_mem_usage = False
1000
            logger.warning(
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
                "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`."
            )

1013
1014
        # 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"):
1015
1016
1017
1018
            raise NotImplementedError(
                "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
                " `device_map=None`."
            )
1019

1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
        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`."
            )
1031

1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
        # 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.")

1062
1063
1064
1065
1066
        user_agent = {
            "diffusers": __version__,
            "file_type": "model",
            "framework": "pytorch",
        }
1067
1068
1069
1070
        unused_kwargs = {}

        # Load config if we don't provide a configuration
        config_path = pretrained_model_name_or_path
1071

1072
1073
1074
1075
1076
1077
1078
1079
1080
        # 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,
1081
            token=token,
1082
1083
1084
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
Marc Sun's avatar
Marc Sun committed
1085
            dduf_entries=dduf_entries,
1086
1087
            **kwargs,
        )
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
        # 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:
            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)
1110
            device_map = hf_quantizer.update_device_map(device_map)
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122

            # 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
1123
1124
        use_keep_in_fp32_modules = cls._keep_in_fp32_modules is not None and (
            hf_quantizer is None or getattr(hf_quantizer, "use_keep_in_fp32_modules", False)
1125
        )
1126

1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
        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 = []
1139

1140
        is_sharded = False
1141
1142
1143
1144
        resolved_model_file = None

        # Determine if we're loading from a directory of sharded checkpoints.
        sharded_metadata = None
1145
1146
        index_file = None
        is_local = os.path.isdir(pretrained_model_name_or_path)
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
        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
1161
            "dduf_entries": dduf_entries,
1162
1163
1164
1165
1166
1167
        }
        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
1168
        if index_file is not None and (dduf_entries or index_file.is_file()):
1169
1170
1171
1172
1173
            is_sharded = True

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

1174
        # load model
1175
        if from_flax:
1176
            resolved_model_file = _get_model_file(
1177
                pretrained_model_name_or_path,
1178
                weights_name=FLAX_WEIGHTS_NAME,
1179
1180
1181
1182
                cache_dir=cache_dir,
                force_download=force_download,
                proxies=proxies,
                local_files_only=local_files_only,
1183
                token=token,
1184
1185
1186
                revision=revision,
                subfolder=subfolder,
                user_agent=user_agent,
1187
                commit_hash=commit_hash,
1188
1189
            )
            model = cls.from_config(config, **unused_kwargs)
1190

1191
1192
1193
            # Convert the weights
            from .modeling_pytorch_flax_utils import load_flax_checkpoint_in_pytorch_model

1194
            model = load_flax_checkpoint_in_pytorch_model(model, resolved_model_file)
1195
        else:
Marc Sun's avatar
Marc Sun committed
1196
            # in the case it is sharded, we have already the index
1197
            if is_sharded:
1198
                resolved_model_file, sharded_metadata = _get_checkpoint_shard_files(
1199
1200
1201
1202
1203
1204
1205
1206
1207
                    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
1208
                    dduf_entries=dduf_entries,
1209
                )
1210
            elif use_safetensors:
1211
                try:
1212
                    resolved_model_file = _get_model_file(
1213
                        pretrained_model_name_or_path,
1214
                        weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant),
1215
1216
1217
1218
                        cache_dir=cache_dir,
                        force_download=force_download,
                        proxies=proxies,
                        local_files_only=local_files_only,
1219
                        token=token,
1220
1221
1222
                        revision=revision,
                        subfolder=subfolder,
                        user_agent=user_agent,
1223
                        commit_hash=commit_hash,
Marc Sun's avatar
Marc Sun committed
1224
                        dduf_entries=dduf_entries,
1225
                    )
1226

1227
                except IOError as e:
1228
                    logger.error(f"An error occurred while trying to fetch {pretrained_model_name_or_path}: {e}")
1229
                    if not allow_pickle:
1230
1231
1232
1233
1234
                        raise
                    logger.warning(
                        "Defaulting to unsafe serialization. Pass `allow_pickle=False` to raise an error instead."
                    )

1235
1236
            if resolved_model_file is None and not is_sharded:
                resolved_model_file = _get_model_file(
1237
                    pretrained_model_name_or_path,
1238
                    weights_name=_add_variant(WEIGHTS_NAME, variant),
1239
1240
1241
1242
                    cache_dir=cache_dir,
                    force_download=force_download,
                    proxies=proxies,
                    local_files_only=local_files_only,
1243
                    token=token,
1244
1245
1246
                    revision=revision,
                    subfolder=subfolder,
                    user_agent=user_agent,
1247
                    commit_hash=commit_hash,
Marc Sun's avatar
Marc Sun committed
1248
                    dduf_entries=dduf_entries,
1249
1250
                )

1251
1252
        if not isinstance(resolved_model_file, list):
            resolved_model_file = [resolved_model_file]
1253

1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
        # set dtype to instantiate the model under:
        # 1. If torch_dtype is not None, we use that dtype
        # 2. If torch_dtype is float8, we don't use _set_default_torch_dtype and we downcast after loading the model
        dtype_orig = None
        if torch_dtype is not None and not torch_dtype == getattr(torch, "float8_e4m3fn", None):
            if 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)}."
                )
            dtype_orig = cls._set_default_torch_dtype(torch_dtype)
1264

1265
        init_contexts = [no_init_weights()]
1266

1267
1268
        if low_cpu_mem_usage:
            init_contexts.append(accelerate.init_empty_weights())
1269

1270
1271
        with ContextManagers(init_contexts):
            model = cls.from_config(config, **unused_kwargs)
1272

1273
1274
        if dtype_orig is not None:
            torch.set_default_dtype(dtype_orig)
1275

1276
1277
1278
1279
1280
1281
        state_dict = None
        if not is_sharded:
            # Time to load the checkpoint
            state_dict = load_state_dict(resolved_model_file[0], disable_mmap=disable_mmap, dduf_entries=dduf_entries)
            # We only fix it for non sharded checkpoints as we don't need it yet for sharded one.
            model._fix_state_dict_keys_on_load(state_dict)
1282

1283
1284
1285
1286
        if is_sharded:
            loaded_keys = sharded_metadata["all_checkpoint_keys"]
        else:
            loaded_keys = list(state_dict.keys())
1287

1288
1289
1290
1291
        if hf_quantizer is not None:
            hf_quantizer.preprocess_model(
                model=model, device_map=device_map, keep_in_fp32_modules=keep_in_fp32_modules
            )
Aryan's avatar
Aryan committed
1292

1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
        # Now that the model is loaded, we can determine the device_map
        device_map = _determine_device_map(
            model, device_map, max_memory, torch_dtype, keep_in_fp32_modules, hf_quantizer
        )
        if hf_quantizer is not None:
            hf_quantizer.validate_environment(device_map=device_map)

        (
            model,
            missing_keys,
            unexpected_keys,
            mismatched_keys,
            offload_index,
            error_msgs,
        ) = cls._load_pretrained_model(
            model,
            state_dict,
            resolved_model_file,
            pretrained_model_name_or_path,
            loaded_keys,
            ignore_mismatched_sizes=ignore_mismatched_sizes,
            low_cpu_mem_usage=low_cpu_mem_usage,
            device_map=device_map,
            offload_folder=offload_folder,
            offload_state_dict=offload_state_dict,
            dtype=torch_dtype,
            hf_quantizer=hf_quantizer,
            keep_in_fp32_modules=keep_in_fp32_modules,
            dduf_entries=dduf_entries,
        )
        loading_info = {
            "missing_keys": missing_keys,
            "unexpected_keys": unexpected_keys,
            "mismatched_keys": mismatched_keys,
            "error_msgs": error_msgs,
        }
1329

1330
1331
1332
1333
1334
1335
1336
1337
        # Dispatch model with hooks on all devices if necessary
        if device_map is not None:
            device_map_kwargs = {
                "device_map": device_map,
                "offload_dir": offload_folder,
                "offload_index": offload_index,
            }
            dispatch_model(model, **device_map_kwargs)
1338

1339
1340
1341
1342
        if hf_quantizer is not None:
            hf_quantizer.postprocess_model(model)
            model.hf_quantizer = hf_quantizer

1343
1344
1345
1346
1347
1348
        if (
            torch_dtype is not None
            and torch_dtype == getattr(torch, "float8_e4m3fn", None)
            and hf_quantizer is None
            and not use_keep_in_fp32_modules
        ):
1349
1350
            model = model.to(torch_dtype)

1351
1352
1353
1354
1355
1356
        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)
1357
1358
1359

        # Set model in evaluation mode to deactivate DropOut modules by default
        model.eval()
1360

1361
        if output_loading_info:
1362
1363
1364
1365
            return model, loading_info

        return model

1366
1367
1368
    # Adapted from `transformers`.
    @wraps(torch.nn.Module.cuda)
    def cuda(self, *args, **kwargs):
Aryan's avatar
Aryan committed
1369
1370
        from ..hooks.group_offloading import _is_group_offload_enabled

1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
        # 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."
                )
Aryan's avatar
Aryan committed
1383
1384
1385
1386
1387
1388
1389
1390

        # Checks if group offloading is enabled
        if _is_group_offload_enabled(self):
            logger.warning(
                f"The module '{self.__class__.__name__}' is group offloaded and moving it using `.cuda()` is not supported."
            )
            return self

1391
1392
1393
1394
1395
        return super().cuda(*args, **kwargs)

    # Adapted from `transformers`.
    @wraps(torch.nn.Module.to)
    def to(self, *args, **kwargs):
Aryan's avatar
Aryan committed
1396
1397
1398
        from ..hooks.group_offloading import _is_group_offload_enabled

        device_arg_or_kwarg_present = any(isinstance(arg, torch.device) for arg in args) or "device" in kwargs
1399
1400
        dtype_present_in_args = "dtype" in kwargs

Aryan's avatar
Aryan committed
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
        # Try converting arguments to torch.device in case they are passed as strings
        for arg in args:
            if not isinstance(arg, str):
                continue
            try:
                torch.device(arg)
                device_arg_or_kwarg_present = True
            except RuntimeError:
                pass

1411
1412
1413
1414
1415
1416
        if not dtype_present_in_args:
            for arg in args:
                if isinstance(arg, torch.dtype):
                    dtype_present_in_args = True
                    break

1417
        if getattr(self, "is_quantized", False):
1418
1419
            if dtype_present_in_args:
                raise ValueError(
1420
1421
                    "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`"
1422
1423
                )

1424
        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
            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."
                )
Aryan's avatar
Aryan committed
1435
1436
1437
1438
1439
1440
1441

        if _is_group_offload_enabled(self) and device_arg_or_kwarg_present:
            logger.warning(
                f"The module '{self.__class__.__name__}' is group offloaded and moving it using `.to()` is not supported."
            )
            return self

1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
        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)

1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
    def compile_repeated_blocks(self, *args, **kwargs):
        """
        Compiles *only* the frequently repeated sub-modules of a model (e.g. the Transformer layers) instead of
        compiling the entire model. This technique—often called **regional compilation** (see the PyTorch recipe
        https://docs.pytorch.org/tutorials/recipes/regional_compilation.html) can reduce end-to-end compile time
        substantially, while preserving the runtime speed-ups you would expect from a full `torch.compile`.

        The set of sub-modules to compile is discovered by the presence of **`_repeated_blocks`** attribute in the
        model definition. Define this attribute on your model subclass as a list/tuple of class names (strings). Every
        module whose class name matches will be compiled.

        Once discovered, each matching sub-module is compiled by calling `submodule.compile(*args, **kwargs)`. Any
        positional or keyword arguments you supply to `compile_repeated_blocks` are forwarded verbatim to
        `torch.compile`.
        """
        repeated_blocks = getattr(self, "_repeated_blocks", None)

        if not repeated_blocks:
            raise ValueError(
                "`_repeated_blocks` attribute is empty. "
                f"Set `_repeated_blocks` for the class `{self.__class__.__name__}` to benefit from faster compilation. "
            )
        has_compiled_region = False
        for submod in self.modules():
            if submod.__class__.__name__ in repeated_blocks:
                submod.compile(*args, **kwargs)
                has_compiled_region = True

        if not has_compiled_region:
            raise ValueError(
                f"Regional compilation failed because {repeated_blocks} classes are not found in the model. "
            )

1499
1500
1501
1502
    @classmethod
    def _load_pretrained_model(
        cls,
        model,
1503
        state_dict: OrderedDict,
1504
        resolved_model_file: List[str],
1505
        pretrained_model_name_or_path: Union[str, os.PathLike],
1506
        loaded_keys: List[str],
1507
        ignore_mismatched_sizes: bool = False,
1508
1509
1510
1511
1512
        assign_to_params_buffers: bool = False,
        hf_quantizer: Optional[DiffusersQuantizer] = None,
        low_cpu_mem_usage: bool = True,
        dtype: Optional[Union[str, torch.dtype]] = None,
        keep_in_fp32_modules: Optional[List[str]] = None,
1513
        device_map: Union[str, int, torch.device, Dict[str, Union[int, str, torch.device]]] = None,
1514
1515
1516
        offload_state_dict: Optional[bool] = None,
        offload_folder: Optional[Union[str, os.PathLike]] = None,
        dduf_entries: Optional[Dict[str, DDUFEntry]] = None,
1517
1518
1519
1520
    ):
        model_state_dict = model.state_dict()
        expected_keys = list(model_state_dict.keys())
        missing_keys = list(set(expected_keys) - set(loaded_keys))
1521
1522
        if hf_quantizer is not None:
            missing_keys = hf_quantizer.update_missing_keys(model, missing_keys, prefix="")
1523
        unexpected_keys = list(set(loaded_keys) - set(expected_keys))
1524
1525
1526
1527
1528
        # Some models may have keys that are not in the state by design, removing them before needlessly warning
        # the user.
        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]
1529

1530
1531
1532
1533
1534
1535
1536
1537
        # Deal with offload
        if device_map is not None and "disk" in device_map.values():
            if offload_folder is None:
                raise ValueError(
                    "The current `device_map` had weights offloaded to the disk. Please provide an `offload_folder`"
                    " for them. Alternatively, make sure you have `safetensors` installed if the model you are using"
                    " offers the weights in this format."
                )
1538
            else:
1539
1540
1541
1542
                os.makedirs(offload_folder, exist_ok=True)
            if offload_state_dict is None:
                offload_state_dict = True

1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
        # If a device map has been used, we can speedup the load time by warming up the device caching allocator.
        # If we don't warmup, each tensor allocation on device calls to the allocator for memory (effectively, a
        # lot of individual calls to device malloc). We can, however, preallocate the memory required by the
        # tensors using their expected shape and not performing any initialization of the memory (empty data).
        # When the actual device allocations happen, the allocator already has a pool of unused device memory
        # that it can re-use for faster loading of the model.
        # TODO: add support for warmup with hf_quantizer
        if device_map is not None and hf_quantizer is None:
            expanded_device_map = _expand_device_map(device_map, expected_keys)
            _caching_allocator_warmup(model, expanded_device_map, dtype)

1554
        offload_index = {} if device_map is not None and "disk" in device_map.values() else None
1555
        state_dict_folder, state_dict_index = None, None
1556
1557
1558
        if offload_state_dict:
            state_dict_folder = tempfile.mkdtemp()
            state_dict_index = {}
1559
1560

        if state_dict is not None:
1561
1562
1563
1564
1565
1566
1567
            # load_state_dict will manage the case where we pass a dict instead of a file
            # if state dict is not None, it means that we don't need to read the files from resolved_model_file also
            resolved_model_file = [state_dict]

        if len(resolved_model_file) > 1:
            resolved_model_file = logging.tqdm(resolved_model_file, desc="Loading checkpoint shards")

1568
1569
1570
1571
        mismatched_keys = []
        assign_to_params_buffers = None
        error_msgs = []

1572
1573
1574
        for shard_file in resolved_model_file:
            state_dict = load_state_dict(shard_file, dduf_entries=dduf_entries)
            mismatched_keys += _find_mismatched_keys(
1575
                state_dict, model_state_dict, loaded_keys, ignore_mismatched_sizes
1576
            )
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596

            if low_cpu_mem_usage:
                offload_index, state_dict_index = load_model_dict_into_meta(
                    model,
                    state_dict,
                    device_map=device_map,
                    dtype=dtype,
                    hf_quantizer=hf_quantizer,
                    keep_in_fp32_modules=keep_in_fp32_modules,
                    unexpected_keys=unexpected_keys,
                    offload_folder=offload_folder,
                    offload_index=offload_index,
                    state_dict_index=state_dict_index,
                    state_dict_folder=state_dict_folder,
                )
            else:
                if assign_to_params_buffers is None:
                    assign_to_params_buffers = check_support_param_buffer_assignment(model, state_dict)
                error_msgs += _load_state_dict_into_model(model, state_dict, assign_to_params_buffers)

1597
1598
        empty_device_cache()

1599
1600
1601
1602
1603
1604
1605
        if offload_index is not None and len(offload_index) > 0:
            save_offload_index(offload_index, offload_folder)
            offload_index = None

            if offload_state_dict:
                load_offloaded_weights(model, state_dict_index, state_dict_folder)
                shutil.rmtree(state_dict_folder)
1606
1607
1608
1609
1610
1611
1612
1613
1614

        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
1615
1616
        if len(unexpected_keys) > 0:
            logger.warning(
1617
                f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
Patrick von Platen's avatar
Patrick von Platen committed
1618
1619
1620
            )
        else:
            logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
1621

Patrick von Platen's avatar
Patrick von Platen committed
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
        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."
            )
1648

1649
        return model, missing_keys, unexpected_keys, mismatched_keys, offload_index, error_msgs
1650

1651
1652
1653
1654
1655
1656
1657
1658
1659
    @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

1660
1661
1662
    # Adapted from `transformers` modeling_utils.py
    def _get_no_split_modules(self, device_map: str):
        """
1663
        Get the modules of the model that should not be split when using device_map. We iterate through the modules to
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
        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)

1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
    @classmethod
    def _set_default_torch_dtype(cls, dtype: torch.dtype) -> torch.dtype:
        """
        Change the default dtype and return the previous one. This is needed when wanting to instantiate the model
        under specific dtype.

        Args:
            dtype (`torch.dtype`):
                a floating dtype to set to.

        Returns:
            `torch.dtype`: the original `dtype` that can be used to restore `torch.set_default_dtype(dtype)` if it was
            modified. If it wasn't, returns `None`.

        Note `set_default_dtype` currently only works with floating-point types and asserts if for example,
        `torch.int64` is passed. So if a non-float `dtype` is passed this functions will throw an exception.
        """
        if not dtype.is_floating_point:
            raise ValueError(
                f"Can't instantiate {cls.__name__} model under dtype={dtype} since it is not a floating point dtype"
            )

        logger.info(f"Instantiating {cls.__name__} model under default dtype {dtype}.")
        dtype_orig = torch.get_default_dtype()
        torch.set_default_dtype(dtype)
        return dtype_orig

1717
    @property
1718
    def device(self) -> torch.device:
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
        """
        `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
1734
        Get number of (trainable or non-embedding) parameters in the module.
1735
1736
1737

        Args:
            only_trainable (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
1738
                Whether or not to return only the number of trainable parameters.
1739
            exclude_embeddings (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
1740
                Whether or not to return only the number of non-embedding parameters.
1741
1742
1743

        Returns:
            `int`: The number of parameters.
Steven Liu's avatar
Steven Liu committed
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754

        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
        ```
1755
        """
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
        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. "
                )
1766
1767
1768

        if exclude_embeddings:
            embedding_param_names = [
1769
                f"{name}.weight" for name, module_type in self.named_modules() if isinstance(module_type, nn.Embedding)
1770
            ]
1771
            total_parameters = [
1772
1773
1774
                parameter for name, parameter in self.named_parameters() if name not in embedding_param_names
            ]
        else:
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
            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
1813

1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
    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`."
            )

1832
1833
1834
1835
1836
1837
1838
    def _fix_state_dict_keys_on_load(self, state_dict: OrderedDict) -> None:
        """
        This function fix the state dict of the model to take into account some changes that were made in the model
        architecture:
        - deprecated attention blocks (happened before we introduced sharded checkpoint,
        so this is why we apply this method only when loading non sharded checkpoints for now)
        """
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
        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")
1881
        return state_dict
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892


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):
1893
        # To prevent dependency import problem.
1894
1895
        from .model_loading_utils import _fetch_remapped_cls_from_config

1896
1897
1898
        # Create a copy of the kwargs so that we don't mess with the keyword arguments in the downstream calls.
        kwargs_copy = kwargs.copy()

1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
        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)

1934
1935
1936
1937
1938
1939
        if remapped_class is cls:
            return super(LegacyModelMixin, remapped_class).from_pretrained(
                pretrained_model_name_or_path, **kwargs_copy
            )
        else:
            return remapped_class.from_pretrained(pretrained_model_name_or_path, **kwargs_copy)