lora_pipeline.py 317 KB
Newer Older
Aryan's avatar
Aryan committed
1
# Copyright 2025 The HuggingFace Team. All rights reserved.
2
3
4
5
6
7
8
9
10
11
12
13
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
Aryan's avatar
Aryan committed
14

15
16
17
18
import os
from typing import Callable, Dict, List, Optional, Union

import torch
19
from huggingface_hub.utils import validate_hf_hub_args
20
21
22

from ..utils import (
    USE_PEFT_BACKEND,
23
    deprecate,
24
    get_submodule_by_name,
hlky's avatar
hlky committed
25
26
    is_bitsandbytes_available,
    is_gguf_available,
27
    is_peft_available,
28
    is_peft_version,
29
    is_torch_version,
30
    is_transformers_available,
31
    is_transformers_version,
32
33
    logging,
)
34
35
36
37
38
39
from .lora_base import (  # noqa
    LORA_WEIGHT_NAME,
    LORA_WEIGHT_NAME_SAFE,
    LoraBaseMixin,
    _fetch_state_dict,
    _load_lora_into_text_encoder,
40
    _pack_dict_with_prefix,
41
)
42
from .lora_conversion_utils import (
Aryan's avatar
Aryan committed
43
    _convert_bfl_flux_control_lora_to_diffusers,
44
    _convert_fal_kontext_lora_to_diffusers,
45
    _convert_hunyuan_video_lora_to_diffusers,
46
    _convert_kohya_flux_lora_to_diffusers,
47
    _convert_musubi_wan_lora_to_diffusers,
48
    _convert_non_diffusers_hidream_lora_to_diffusers,
49
    _convert_non_diffusers_lora_to_diffusers,
50
    _convert_non_diffusers_ltxv_lora_to_diffusers,
51
    _convert_non_diffusers_lumina2_lora_to_diffusers,
52
    _convert_non_diffusers_wan_lora_to_diffusers,
53
54
55
    _convert_xlabs_flux_lora_to_diffusers,
    _maybe_map_sgm_blocks_to_diffusers,
)
56
57


58
59
60
61
62
63
64
65
66
67
68
_LOW_CPU_MEM_USAGE_DEFAULT_LORA = False
if is_torch_version(">=", "1.9.0"):
    if (
        is_peft_available()
        and is_peft_version(">=", "0.13.1")
        and is_transformers_available()
        and is_transformers_version(">", "4.45.2")
    ):
        _LOW_CPU_MEM_USAGE_DEFAULT_LORA = True


69
70
71
72
logger = logging.get_logger(__name__)

TEXT_ENCODER_NAME = "text_encoder"
UNET_NAME = "unet"
Will Berman's avatar
Will Berman committed
73
TRANSFORMER_NAME = "transformer"
74

Aryan's avatar
Aryan committed
75
76
_MODULE_NAME_TO_ATTRIBUTE_MAP_FLUX = {"x_embedder": "in_channels"}

77

hlky's avatar
hlky committed
78
79
80
81
82
83
84
85
def _maybe_dequantize_weight_for_expanded_lora(model, module):
    if is_bitsandbytes_available():
        from ..quantizers.bitsandbytes import dequantize_bnb_weight

    if is_gguf_available():
        from ..quantizers.gguf.utils import dequantize_gguf_tensor

    is_bnb_4bit_quantized = module.weight.__class__.__name__ == "Params4bit"
86
    is_bnb_8bit_quantized = module.weight.__class__.__name__ == "Int8Params"
hlky's avatar
hlky committed
87
88
89
90
91
92
    is_gguf_quantized = module.weight.__class__.__name__ == "GGUFParameter"

    if is_bnb_4bit_quantized and not is_bitsandbytes_available():
        raise ValueError(
            "The checkpoint seems to have been quantized with `bitsandbytes` (4bits). Install `bitsandbytes` to load quantized checkpoints."
        )
93
94
95
96
    if is_bnb_8bit_quantized and not is_bitsandbytes_available():
        raise ValueError(
            "The checkpoint seems to have been quantized with `bitsandbytes` (8bits). Install `bitsandbytes` to load quantized checkpoints."
        )
hlky's avatar
hlky committed
97
98
99
100
101
102
    if is_gguf_quantized and not is_gguf_available():
        raise ValueError(
            "The checkpoint seems to have been quantized with `gguf`. Install `gguf` to load quantized checkpoints."
        )

    weight_on_cpu = False
103
    if module.weight.device.type == "cpu":
hlky's avatar
hlky committed
104
105
        weight_on_cpu = True

106
    device = torch.accelerator.current_accelerator().type if hasattr(torch, "accelerator") else "cuda"
107
    if is_bnb_4bit_quantized or is_bnb_8bit_quantized:
hlky's avatar
hlky committed
108
        module_weight = dequantize_bnb_weight(
109
            module.weight.to(device) if weight_on_cpu else module.weight,
110
            state=module.weight.quant_state if is_bnb_4bit_quantized else module.state,
hlky's avatar
hlky committed
111
112
113
114
            dtype=model.dtype,
        ).data
    elif is_gguf_quantized:
        module_weight = dequantize_gguf_tensor(
115
            module.weight.to(device) if weight_on_cpu else module.weight,
hlky's avatar
hlky committed
116
117
118
119
120
121
122
123
124
125
126
        )
        module_weight = module_weight.to(model.dtype)
    else:
        module_weight = module.weight.data

    if weight_on_cpu:
        module_weight = module_weight.cpu()

    return module_weight


127
class StableDiffusionLoraLoaderMixin(LoraBaseMixin):
128
    r"""
129
    Load LoRA layers into Stable Diffusion [`UNet2DConditionModel`] and
130
    [`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel).
131
    """
132

133
    _lora_loadable_modules = ["unet", "text_encoder"]
134
    unet_name = UNET_NAME
135
    text_encoder_name = TEXT_ENCODER_NAME
136
137

    def load_lora_weights(
138
139
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
140
        adapter_name: Optional[str] = None,
141
142
        hotswap: bool = False,
        **kwargs,
143
    ):
144
        """Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
145
146
147
148
        `self.text_encoder`.

        All kwargs are forwarded to `self.lora_state_dict`.

149
150
        See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is
        loaded.
151

152
153
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet`] for more details on how the state dict is
        loaded into `self.unet`.
154

155
156
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder`] for more details on how the state
        dict is loaded into `self.text_encoder`.
157
158
159

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
160
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
161
            adapter_name (`str`, *optional*):
162
163
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
164
165
166
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
167
            hotswap (`bool`, *optional*):
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
                Defaults to `False`. Whether to substitute an existing (LoRA) adapter with the newly loaded adapter
                in-place. This means that, instead of loading an additional adapter, this will take the existing
                adapter weights and replace them with the weights of the new adapter. This can be faster and more
                memory efficient. However, the main advantage of hotswapping is that when the model is compiled with
                torch.compile, loading the new adapter does not require recompilation of the model. When using
                hotswapping, the passed `adapter_name` should be the name of an already loaded adapter.

                If the new adapter and the old adapter have different ranks and/or LoRA alphas (i.e. scaling), you need
                to call an additional method before loading the adapter:

                ```py
                pipeline = ...  # load diffusers pipeline
                max_rank = ...  # the highest rank among all LoRAs that you want to load
                # call *before* compiling and loading the LoRA adapter
                pipeline.enable_lora_hotswap(target_rank=max_rank)
                pipeline.load_lora_weights(file_name)
                # optionally compile the model now
                ```

                Note that hotswapping adapters of the text encoder is not yet supported. There are some further
                limitations to this technique, which are documented here:
                https://huggingface.co/docs/peft/main/en/package_reference/hotswap
190
191
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
192
        """
193
194
195
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

196
197
198
199
200
201
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and not is_peft_version(">=", "0.13.1"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

202
203
204
205
        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

206
        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
207
208
        kwargs["return_lora_metadata"] = True
        state_dict, network_alphas, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
209

Sayak Paul's avatar
Sayak Paul committed
210
        is_correct_format = all("lora" in key for key in state_dict.keys())
211
212
213
214
215
216
217
218
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_unet(
            state_dict,
            network_alphas=network_alphas,
            unet=getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet,
            adapter_name=adapter_name,
219
            metadata=metadata,
220
            _pipeline=self,
221
            low_cpu_mem_usage=low_cpu_mem_usage,
222
            hotswap=hotswap,
223
224
225
226
227
228
229
230
231
232
        )
        self.load_lora_into_text_encoder(
            state_dict,
            network_alphas=network_alphas,
            text_encoder=getattr(self, self.text_encoder_name)
            if not hasattr(self, "text_encoder")
            else self.text_encoder,
            lora_scale=self.lora_scale,
            adapter_name=adapter_name,
            _pipeline=self,
233
            metadata=metadata,
234
            low_cpu_mem_usage=low_cpu_mem_usage,
235
            hotswap=hotswap,
236
237
238
        )

    @classmethod
239
    @validate_hf_hub_args
240
241
242
243
244
245
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
246
247
248
249
250
251
252
253
254
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>
255
256
257
258
259
260
261

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
                      the Hub.
262
263
                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
                      with [`ModelMixin.save_pretrained`].
264
265
266
267
268
269
270
271
272
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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.
273

274
275
276
277
278
279
            proxies (`Dict[str, str]`, *optional*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
280
            token (`str` or *bool*, *optional*):
281
282
283
284
285
286
287
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
288
289
            weight_name (`str`, *optional*, defaults to None):
                Name of the serialized state dict file.
290
291
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
292
293
294
        """
        # Load the main state dict first which has the LoRA layers for either of
        # UNet and text encoder or both.
295
        cache_dir = kwargs.pop("cache_dir", None)
296
297
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
298
299
        local_files_only = kwargs.pop("local_files_only", None)
        token = kwargs.pop("token", None)
300
301
302
303
304
        revision = kwargs.pop("revision", None)
        subfolder = kwargs.pop("subfolder", None)
        weight_name = kwargs.pop("weight_name", None)
        unet_config = kwargs.pop("unet_config", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
305
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
306
307
308
309
310
311

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

312
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
313

314
        state_dict, metadata = _fetch_state_dict(
315
316
317
318
319
320
321
322
323
324
325
326
327
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )
Sayak Paul's avatar
Sayak Paul committed
328
329
330
331
332
        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347

        network_alphas = None
        # TODO: replace it with a method from `state_dict_utils`
        if all(
            (
                k.startswith("lora_te_")
                or k.startswith("lora_unet_")
                or k.startswith("lora_te1_")
                or k.startswith("lora_te2_")
            )
            for k in state_dict.keys()
        ):
            # Map SDXL blocks correctly.
            if unet_config is not None:
                # use unet config to remap block numbers
348
                state_dict = _maybe_map_sgm_blocks_to_diffusers(state_dict, unet_config)
349
            state_dict, network_alphas = _convert_non_diffusers_lora_to_diffusers(state_dict)
350

351
352
        out = (state_dict, network_alphas, metadata) if return_lora_metadata else (state_dict, network_alphas)
        return out
353
354

    @classmethod
355
    def load_lora_into_unet(
356
357
358
359
360
361
362
363
        cls,
        state_dict,
        network_alphas,
        unet,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
364
        metadata=None,
365
    ):
366
        """
367
        This will load the LoRA layers specified in `state_dict` into `unet`.
368
369
370

        Parameters:
            state_dict (`dict`):
371
372
373
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
374
            network_alphas (`Dict[str, float]`):
375
376
377
                The value of the network alpha used for stable learning and preventing underflow. This value has the
                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
378
379
380
            unet (`UNet2DConditionModel`):
                The UNet model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
381
382
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
383
384
385
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading only loading the pretrained LoRA weights and not initializing the random
                weights.
386
387
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
388
389
390
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
391
        """
392
393
394
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

395
396
397
398
399
        if low_cpu_mem_usage and not is_peft_version(">=", "0.13.1"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

400
401
402
        # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
        # then the `state_dict` keys should have `cls.unet_name` and/or `cls.text_encoder_name` as
        # their prefixes.
403
404
405
406
407
408
        logger.info(f"Loading {cls.unet_name}.")
        unet.load_lora_adapter(
            state_dict,
            prefix=cls.unet_name,
            network_alphas=network_alphas,
            adapter_name=adapter_name,
409
            metadata=metadata,
410
411
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
412
            hotswap=hotswap,
413
        )
414

415
416
417
418
419
420
421
422
423
424
    @classmethod
    def load_lora_into_text_encoder(
        cls,
        state_dict,
        network_alphas,
        text_encoder,
        prefix=None,
        lora_scale=1.0,
        adapter_name=None,
        _pipeline=None,
425
        low_cpu_mem_usage=False,
426
        hotswap: bool = False,
427
        metadata=None,
428
429
430
431
432
433
434
435
436
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `text_encoder`

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The key should be prefixed with an
                additional `text_encoder` to distinguish between unet lora layers.
            network_alphas (`Dict[str, float]`):
437
438
439
                The value of the network alpha used for stable learning and preventing underflow. This value has the
                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
440
441
442
443
444
445
446
447
448
449
            text_encoder (`CLIPTextModel`):
                The text encoder model to load the LoRA layers into.
            prefix (`str`):
                Expected prefix of the `text_encoder` in the `state_dict`.
            lora_scale (`float`):
                How much to scale the output of the lora linear layer before it is added with the output of the regular
                lora layer.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
450
451
452
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
453
454
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
455
456
457
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
458
        """
459
460
461
462
463
464
465
466
        _load_lora_into_text_encoder(
            state_dict=state_dict,
            network_alphas=network_alphas,
            lora_scale=lora_scale,
            text_encoder=text_encoder,
            prefix=prefix,
            text_encoder_name=cls.text_encoder_name,
            adapter_name=adapter_name,
467
            metadata=metadata,
468
469
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
470
            hotswap=hotswap,
471
        )
472

473
474
475
476
477
478
479
480
481
482
    @classmethod
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        text_encoder_lora_layers: Dict[str, torch.nn.Module] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
483
484
        unet_lora_adapter_metadata=None,
        text_encoder_lora_adapter_metadata=None,
485
486
    ):
        r"""
487
        Save the LoRA parameters corresponding to the UNet and text encoder.
488
489
490

        Arguments:
            save_directory (`str` or `os.PathLike`):
491
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
492
493
494
495
496
497
498
499
500
501
            unet_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `unet`.
            text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
                encoder LoRA state dict because it comes from 🤗 Transformers.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
502
503
                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
504
505
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
506
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
507
508
509
510
            unet_lora_adapter_metadata:
                LoRA adapter metadata associated with the unet to be serialized with the state dict.
            text_encoder_lora_adapter_metadata:
                LoRA adapter metadata associated with the text encoder to be serialized with the state dict.
511
512
        """
        state_dict = {}
513
        lora_adapter_metadata = {}
514

515
516
        if not (unet_lora_layers or text_encoder_lora_layers):
            raise ValueError("You must pass at least one of `unet_lora_layers` and `text_encoder_lora_layers`.")
517

518
        if unet_lora_layers:
519
            state_dict.update(cls.pack_weights(unet_lora_layers, cls.unet_name))
520

521
        if text_encoder_lora_layers:
522
            state_dict.update(cls.pack_weights(text_encoder_lora_layers, cls.text_encoder_name))
Will Berman's avatar
Will Berman committed
523

524
525
526
527
528
529
530
531
        if unet_lora_adapter_metadata:
            lora_adapter_metadata.update(_pack_dict_with_prefix(unet_lora_adapter_metadata, cls.unet_name))

        if text_encoder_lora_adapter_metadata:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(text_encoder_lora_adapter_metadata, cls.text_encoder_name)
            )

532
533
534
535
536
537
538
539
        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
540
            lora_adapter_metadata=lora_adapter_metadata,
541
542
        )

543
544
    def fuse_lora(
        self,
545
        components: List[str] = ["unet", "text_encoder"],
546
547
548
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
549
        **kwargs,
550
551
552
553
554
555
556
557
558
559
560
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
561
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
582
        super().fuse_lora(
583
584
585
586
587
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
588
        )
589

590
    def unfuse_lora(self, components: List[str] = ["unet", "text_encoder"], **kwargs):
591
592
        r"""
        Reverses the effect of
593
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).
594
595
596
597
598
599
600
601

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
602
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
603
604
605
606
607
            unfuse_unet (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
            unfuse_text_encoder (`bool`, defaults to `True`):
                Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
                LoRA parameters then it won't have any effect.
        """
608
        super().unfuse_lora(components=components, **kwargs)
609
610


611
612
613
614
615
616
class StableDiffusionXLLoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into Stable Diffusion XL [`UNet2DConditionModel`],
    [`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), and
    [`CLIPTextModelWithProjection`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection).
    """
617

618
619
620
    _lora_loadable_modules = ["unet", "text_encoder", "text_encoder_2"]
    unet_name = UNET_NAME
    text_encoder_name = TEXT_ENCODER_NAME
621

622
623
624
625
    def load_lora_weights(
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
626
        hotswap: bool = False,
627
628
629
630
631
632
633
634
        **kwargs,
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
        `self.text_encoder`.

        All kwargs are forwarded to `self.lora_state_dict`.

635
636
        See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is
        loaded.
637

638
639
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet`] for more details on how the state dict is
        loaded into `self.unet`.
640

641
642
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder`] for more details on how the state
        dict is loaded into `self.text_encoder`.
643
644
645

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
646
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
Steven Liu's avatar
Steven Liu committed
647
            adapter_name (`str`, *optional*):
648
649
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
650
651
652
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
653
654
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
655
            kwargs (`dict`, *optional*):
656
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
657
        """
658
659
660
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

661
662
663
664
665
666
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and not is_peft_version(">=", "0.13.1"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

667
668
669
670
        # We could have accessed the unet config from `lora_state_dict()` too. We pass
        # it here explicitly to be able to tell that it's coming from an SDXL
        # pipeline.

671
672
673
674
        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

675
        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
676
677
        kwargs["return_lora_metadata"] = True
        state_dict, network_alphas, metadata = self.lora_state_dict(
678
679
680
681
            pretrained_model_name_or_path_or_dict,
            unet_config=self.unet.config,
            **kwargs,
        )
Sayak Paul's avatar
Sayak Paul committed
682
683

        is_correct_format = all("lora" in key for key in state_dict.keys())
684
685
686
687
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_unet(
688
689
690
691
            state_dict,
            network_alphas=network_alphas,
            unet=self.unet,
            adapter_name=adapter_name,
692
            metadata=metadata,
693
694
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
695
            hotswap=hotswap,
696
        )
697
698
699
700
701
702
703
        self.load_lora_into_text_encoder(
            state_dict,
            network_alphas=network_alphas,
            text_encoder=self.text_encoder,
            prefix=self.text_encoder_name,
            lora_scale=self.lora_scale,
            adapter_name=adapter_name,
704
            metadata=metadata,
705
706
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
707
            hotswap=hotswap,
708
709
710
711
712
713
714
715
        )
        self.load_lora_into_text_encoder(
            state_dict,
            network_alphas=network_alphas,
            text_encoder=self.text_encoder_2,
            prefix=f"{self.text_encoder_name}_2",
            lora_scale=self.lora_scale,
            adapter_name=adapter_name,
716
            metadata=metadata,
717
718
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
719
            hotswap=hotswap,
720
        )
721
722

    @classmethod
723
724
725
    @validate_hf_hub_args
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.lora_state_dict
    def lora_state_dict(
726
727
728
729
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
730
        r"""
731
        Return state dict for lora weights and the network alphas.
732
733

        <Tip warning={true}>
Dhruv Nair's avatar
Dhruv Nair committed
734

735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
773
774
            weight_name (`str`, *optional*, defaults to None):
                Name of the serialized state dict file.
775
776
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
777
778
779
780
781
782
783
784
785
786
787
        """
        # Load the main state dict first which has the LoRA layers for either of
        # UNet and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
788
        unet_config = kwargs.pop("unet_config", None)
789
        use_safetensors = kwargs.pop("use_safetensors", None)
790
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
Dhruv Nair's avatar
Dhruv Nair committed
791

792
793
794
795
796
        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

797
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
798

799
        state_dict, metadata = _fetch_state_dict(
800
801
802
803
804
805
806
807
808
809
810
811
812
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )
Sayak Paul's avatar
Sayak Paul committed
813
814
815
816
817
        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835

        network_alphas = None
        # TODO: replace it with a method from `state_dict_utils`
        if all(
            (
                k.startswith("lora_te_")
                or k.startswith("lora_unet_")
                or k.startswith("lora_te1_")
                or k.startswith("lora_te2_")
            )
            for k in state_dict.keys()
        ):
            # Map SDXL blocks correctly.
            if unet_config is not None:
                # use unet config to remap block numbers
                state_dict = _maybe_map_sgm_blocks_to_diffusers(state_dict, unet_config)
            state_dict, network_alphas = _convert_non_diffusers_lora_to_diffusers(state_dict)

836
837
        out = (state_dict, network_alphas, metadata) if return_lora_metadata else (state_dict, network_alphas)
        return out
838
839
840

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_unet
841
    def load_lora_into_unet(
842
843
844
845
846
847
848
849
        cls,
        state_dict,
        network_alphas,
        unet,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
850
        metadata=None,
851
    ):
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
        """
        This will load the LoRA layers specified in `state_dict` into `unet`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            network_alphas (`Dict[str, float]`):
                The value of the network alpha used for stable learning and preventing underflow. This value has the
                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
            unet (`UNet2DConditionModel`):
                The UNet model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
869
870
871
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading only loading the pretrained LoRA weights and not initializing the random
                weights.
872
873
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
874
875
876
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
877
878
879
880
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

881
882
883
884
885
        if low_cpu_mem_usage and not is_peft_version(">=", "0.13.1"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

886
887
888
        # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
        # then the `state_dict` keys should have `cls.unet_name` and/or `cls.text_encoder_name` as
        # their prefixes.
889
890
891
892
893
894
        logger.info(f"Loading {cls.unet_name}.")
        unet.load_lora_adapter(
            state_dict,
            prefix=cls.unet_name,
            network_alphas=network_alphas,
            adapter_name=adapter_name,
895
            metadata=metadata,
896
897
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
898
            hotswap=hotswap,
899
        )
900
901
902
903
904
905
906
907
908
909
910
911

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder
    def load_lora_into_text_encoder(
        cls,
        state_dict,
        network_alphas,
        text_encoder,
        prefix=None,
        lora_scale=1.0,
        adapter_name=None,
        _pipeline=None,
912
        low_cpu_mem_usage=False,
913
        hotswap: bool = False,
914
        metadata=None,
915
916
917
918
919
920
921
922
923
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `text_encoder`

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The key should be prefixed with an
                additional `text_encoder` to distinguish between unet lora layers.
            network_alphas (`Dict[str, float]`):
924
925
926
                The value of the network alpha used for stable learning and preventing underflow. This value has the
                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
927
928
929
930
931
932
933
934
935
936
            text_encoder (`CLIPTextModel`):
                The text encoder model to load the LoRA layers into.
            prefix (`str`):
                Expected prefix of the `text_encoder` in the `state_dict`.
            lora_scale (`float`):
                How much to scale the output of the lora linear layer before it is added with the output of the regular
                lora layer.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
937
938
939
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
940
941
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
942
943
944
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
945
        """
946
947
948
949
950
951
952
953
        _load_lora_into_text_encoder(
            state_dict=state_dict,
            network_alphas=network_alphas,
            lora_scale=lora_scale,
            text_encoder=text_encoder,
            prefix=prefix,
            text_encoder_name=cls.text_encoder_name,
            adapter_name=adapter_name,
954
            metadata=metadata,
955
956
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
957
            hotswap=hotswap,
958
        )
959
960
961
962
963
964
965
966
967
968
969
970

    @classmethod
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
971
972
973
        unet_lora_adapter_metadata=None,
        text_encoder_lora_adapter_metadata=None,
        text_encoder_2_lora_adapter_metadata=None,
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
    ):
        r"""
        Save the LoRA parameters corresponding to the UNet and text encoder.

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            unet_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `unet`.
            text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
                encoder LoRA state dict because it comes from 🤗 Transformers.
            text_encoder_2_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `text_encoder_2`. Must explicitly pass the text
                encoder LoRA state dict because it comes from 🤗 Transformers.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
999
1000
1001
1002
1003
1004
            unet_lora_adapter_metadata:
                LoRA adapter metadata associated with the unet to be serialized with the state dict.
            text_encoder_lora_adapter_metadata:
                LoRA adapter metadata associated with the text encoder to be serialized with the state dict.
            text_encoder_2_lora_adapter_metadata:
                LoRA adapter metadata associated with the second text encoder to be serialized with the state dict.
1005
1006
        """
        state_dict = {}
1007
        lora_adapter_metadata = {}
1008
1009
1010

        if not (unet_lora_layers or text_encoder_lora_layers or text_encoder_2_lora_layers):
            raise ValueError(
1011
                "You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers`, `text_encoder_2_lora_layers`."
1012
1013
1014
            )

        if unet_lora_layers:
1015
            state_dict.update(cls.pack_weights(unet_lora_layers, cls.unet_name))
1016
1017
1018
1019
1020
1021
1022

        if text_encoder_lora_layers:
            state_dict.update(cls.pack_weights(text_encoder_lora_layers, "text_encoder"))

        if text_encoder_2_lora_layers:
            state_dict.update(cls.pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))

1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
        if unet_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(_pack_dict_with_prefix(unet_lora_adapter_metadata, cls.unet_name))

        if text_encoder_lora_adapter_metadata:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(text_encoder_lora_adapter_metadata, cls.text_encoder_name)
            )

        if text_encoder_2_lora_adapter_metadata:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(text_encoder_2_lora_adapter_metadata, "text_encoder_2")
            )

1036
1037
1038
1039
1040
1041
1042
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
1043
            lora_adapter_metadata=lora_adapter_metadata,
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
        )

    def fuse_lora(
        self,
        components: List[str] = ["unet", "text_encoder", "text_encoder_2"],
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
1086
1087
1088
1089
1090
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
        )

    def unfuse_lora(self, components: List[str] = ["unet", "text_encoder", "text_encoder_2"], **kwargs):
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_unet (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
            unfuse_text_encoder (`bool`, defaults to `True`):
                Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
                LoRA parameters then it won't have any effect.
        """
1111
        super().unfuse_lora(components=components, **kwargs)
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176


class SD3LoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`SD3Transformer2DModel`],
    [`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), and
    [`CLIPTextModelWithProjection`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection).

    Specific to [`StableDiffusion3Pipeline`].
    """

    _lora_loadable_modules = ["transformer", "text_encoder", "text_encoder_2"]
    transformer_name = TRANSFORMER_NAME
    text_encoder_name = TEXT_ENCODER_NAME

    @classmethod
    @validate_hf_hub_args
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
1177
1178
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191

        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
1192
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
1193
1194
1195
1196
1197
1198

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

1199
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
1200

1201
        state_dict, metadata = _fetch_state_dict(
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

Sayak Paul's avatar
Sayak Paul committed
1216
1217
1218
1219
1220
1221
        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

1222
1223
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
1224
1225

    def load_lora_weights(
1226
1227
1228
1229
1230
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name=None,
        hotswap: bool = False,
        **kwargs,
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
        `self.text_encoder`.

        All kwargs are forwarded to `self.lora_state_dict`.

        See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is
        loaded.

        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
1250
1251
1252
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
1253
1254
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
1255
1256
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
1257
1258
1259
1260
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

1261
1262
1263
1264
1265
1266
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

1267
1268
1269
1270
1271
        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
1272
1273
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
1274

Sayak Paul's avatar
Sayak Paul committed
1275
        is_correct_format = all("lora" in key for key in state_dict.keys())
1276
1277
1278
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

1279
1280
1281
1282
        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
1283
            metadata=metadata,
1284
1285
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
1286
            hotswap=hotswap,
1287
1288
1289
1290
1291
1292
1293
1294
        )
        self.load_lora_into_text_encoder(
            state_dict,
            network_alphas=None,
            text_encoder=self.text_encoder,
            prefix=self.text_encoder_name,
            lora_scale=self.lora_scale,
            adapter_name=adapter_name,
1295
            metadata=metadata,
1296
1297
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
1298
            hotswap=hotswap,
1299
1300
1301
1302
1303
1304
1305
1306
        )
        self.load_lora_into_text_encoder(
            state_dict,
            network_alphas=None,
            text_encoder=self.text_encoder_2,
            prefix=f"{self.text_encoder_name}_2",
            lora_scale=self.lora_scale,
            adapter_name=adapter_name,
1307
            metadata=metadata,
1308
1309
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
1310
            hotswap=hotswap,
1311
        )
1312
1313

    @classmethod
1314
    def load_lora_into_transformer(
1315
1316
1317
1318
1319
1320
1321
1322
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
1323
    ):
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`SD3Transformer2DModel`):
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
1337
1338
1339
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
1340
1341
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
1342
1343
1344
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
1345
        """
1346
1347
1348
1349
1350
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

1351
1352
1353
1354
1355
1356
        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
1357
            metadata=metadata,
1358
1359
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
1360
            hotswap=hotswap,
1361
        )
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder
    def load_lora_into_text_encoder(
        cls,
        state_dict,
        network_alphas,
        text_encoder,
        prefix=None,
        lora_scale=1.0,
        adapter_name=None,
        _pipeline=None,
1374
        low_cpu_mem_usage=False,
1375
        hotswap: bool = False,
1376
        metadata=None,
1377
1378
1379
1380
1381
1382
1383
1384
1385
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `text_encoder`

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The key should be prefixed with an
                additional `text_encoder` to distinguish between unet lora layers.
            network_alphas (`Dict[str, float]`):
1386
1387
1388
                The value of the network alpha used for stable learning and preventing underflow. This value has the
                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
            text_encoder (`CLIPTextModel`):
                The text encoder model to load the LoRA layers into.
            prefix (`str`):
                Expected prefix of the `text_encoder` in the `state_dict`.
            lora_scale (`float`):
                How much to scale the output of the lora linear layer before it is added with the output of the regular
                lora layer.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
1399
1400
1401
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
1402
1403
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
1404
1405
1406
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
1407
        """
1408
1409
1410
1411
1412
1413
1414
1415
        _load_lora_into_text_encoder(
            state_dict=state_dict,
            network_alphas=network_alphas,
            lora_scale=lora_scale,
            text_encoder=text_encoder,
            prefix=prefix,
            text_encoder_name=cls.text_encoder_name,
            adapter_name=adapter_name,
1416
            metadata=metadata,
1417
1418
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
1419
            hotswap=hotswap,
1420
        )
1421
1422

    @classmethod
1423
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.save_lora_weights with unet->transformer
1424
1425
1426
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
1427
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
1428
1429
1430
1431
1432
1433
        text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
1434
1435
1436
        transformer_lora_adapter_metadata=None,
        text_encoder_lora_adapter_metadata=None,
        text_encoder_2_lora_adapter_metadata=None,
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
    ):
        r"""
        Save the LoRA parameters corresponding to the UNet and text encoder.

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
                encoder LoRA state dict because it comes from 🤗 Transformers.
            text_encoder_2_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `text_encoder_2`. Must explicitly pass the text
                encoder LoRA state dict because it comes from 🤗 Transformers.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
1462
1463
1464
1465
1466
1467
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
            text_encoder_lora_adapter_metadata:
                LoRA adapter metadata associated with the text encoder to be serialized with the state dict.
            text_encoder_2_lora_adapter_metadata:
                LoRA adapter metadata associated with the second text encoder to be serialized with the state dict.
1468
1469
        """
        state_dict = {}
1470
        lora_adapter_metadata = {}
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485

        if not (transformer_lora_layers or text_encoder_lora_layers or text_encoder_2_lora_layers):
            raise ValueError(
                "You must pass at least one of `transformer_lora_layers`, `text_encoder_lora_layers`, `text_encoder_2_lora_layers`."
            )

        if transformer_lora_layers:
            state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if text_encoder_lora_layers:
            state_dict.update(cls.pack_weights(text_encoder_lora_layers, "text_encoder"))

        if text_encoder_2_lora_layers:
            state_dict.update(cls.pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))

1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )

        if text_encoder_lora_adapter_metadata:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(text_encoder_lora_adapter_metadata, cls.text_encoder_name)
            )

        if text_encoder_2_lora_adapter_metadata:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(text_encoder_2_lora_adapter_metadata, "text_encoder_2")
            )

1501
1502
1503
1504
1505
1506
1507
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
1508
            lora_adapter_metadata=lora_adapter_metadata,
1509
1510
        )

1511
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.fuse_lora with unet->transformer
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
    def fuse_lora(
        self,
        components: List[str] = ["transformer", "text_encoder", "text_encoder_2"],
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
1552
1553
1554
1555
1556
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
1557
1558
        )

1559
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.unfuse_lora with unet->transformer
1560
1561
1562
1563
1564
1565
    def unfuse_lora(self, components: List[str] = ["transformer", "text_encoder", "text_encoder_2"], **kwargs):
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>
Dhruv Nair's avatar
Dhruv Nair committed
1566

1567
        This is an experimental API.
1568

1569
        </Tip>
1570

1571
1572
        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
1573
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
1574
1575
1576
1577
            unfuse_text_encoder (`bool`, defaults to `True`):
                Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
                LoRA parameters then it won't have any effect.
        """
1578
        super().unfuse_lora(components=components, **kwargs)
1579
1580


1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
class AuraFlowLoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`AuraFlowTransformer2DModel`] Specific to [`AuraFlowPipeline`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.lora_state_dict
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
1640
1641
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654

        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
1655
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
1656
1657
1658
1659
1660
1661

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

1662
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
1663

1664
        state_dict, metadata = _fetch_state_dict(
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

1685
1686
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
1687
1688
1689

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
    def load_lora_weights(
1690
1691
1692
1693
1694
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
1712
1713
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
1731
1732
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
1733
1734
1735
1736
1737
1738
1739
1740
1741

        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
1742
            metadata=metadata,
1743
1744
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
1745
            hotswap=hotswap,
1746
1747
1748
1749
1750
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->AuraFlowTransformer2DModel
    def load_lora_into_transformer(
1751
1752
1753
1754
1755
1756
1757
1758
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`AuraFlowTransformer2DModel`):
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
1776
1777
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
1778
1779
1780
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
1793
            metadata=metadata,
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
            hotswap=hotswap,
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
1809
        transformer_lora_adapter_metadata: Optional[dict] = None,
1810
1811
    ):
        r"""
1812
        Save the LoRA parameters corresponding to the transformer.
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
1829
1830
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
1831
1832
        """
        state_dict = {}
1833
        lora_adapter_metadata = {}
1834
1835
1836
1837

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

1838
1839
1840
1841
1842
1843
        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )
1844
1845
1846
1847
1848
1849
1850
1851
1852

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
1853
            lora_adapter_metadata=lora_adapter_metadata,
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
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
        )

    # Copied from diffusers.loaders.lora_pipeline.SanaLoraLoaderMixin.fuse_lora
    def fuse_lora(
        self,
        components: List[str] = ["transformer"],
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
        )

    # Copied from diffusers.loaders.lora_pipeline.SanaLoraLoaderMixin.unfuse_lora
    def unfuse_lora(self, components: List[str] = ["transformer", "text_encoder"], **kwargs):
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
        super().unfuse_lora(components=components, **kwargs)


Sayak Paul's avatar
Sayak Paul committed
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
class FluxLoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`FluxTransformer2DModel`],
    [`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel).

    Specific to [`StableDiffusion3Pipeline`].
    """

    _lora_loadable_modules = ["transformer", "text_encoder"]
    transformer_name = TRANSFORMER_NAME
    text_encoder_name = TEXT_ENCODER_NAME
Aryan's avatar
Aryan committed
1934
    _control_lora_supported_norm_keys = ["norm_q", "norm_k", "norm_added_q", "norm_added_k"]
Sayak Paul's avatar
Sayak Paul committed
1935
1936
1937
1938
1939
1940

    @classmethod
    @validate_hf_hub_args
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
1941
        return_alphas: bool = False,
Sayak Paul's avatar
Sayak Paul committed
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
1987
1988
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
Sayak Paul's avatar
Sayak Paul committed
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
2001
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
Sayak Paul's avatar
Sayak Paul committed
2002
2003
2004
2005
2006
2007

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

2008
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
Sayak Paul's avatar
Sayak Paul committed
2009

2010
        state_dict, metadata = _fetch_state_dict(
Sayak Paul's avatar
Sayak Paul committed
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )
Sayak Paul's avatar
Sayak Paul committed
2024
2025
2026
2027
2028
        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
Sayak Paul's avatar
Sayak Paul committed
2029

2030
2031
2032
2033
2034
        # TODO (sayakpaul): to a follow-up to clean and try to unify the conditions.
        is_kohya = any(".lora_down.weight" in k for k in state_dict)
        if is_kohya:
            state_dict = _convert_kohya_flux_lora_to_diffusers(state_dict)
            # Kohya already takes care of scaling the LoRA parameters with alpha.
2035
2036
2037
2038
2039
2040
2041
            return cls._prepare_outputs(
                state_dict,
                metadata=metadata,
                alphas=None,
                return_alphas=return_alphas,
                return_metadata=return_lora_metadata,
            )
2042
2043
2044
2045
2046

        is_xlabs = any("processor" in k for k in state_dict)
        if is_xlabs:
            state_dict = _convert_xlabs_flux_lora_to_diffusers(state_dict)
            # xlabs doesn't use `alpha`.
2047
2048
2049
2050
2051
2052
2053
            return cls._prepare_outputs(
                state_dict,
                metadata=metadata,
                alphas=None,
                return_alphas=return_alphas,
                return_metadata=return_lora_metadata,
            )
2054

Aryan's avatar
Aryan committed
2055
2056
2057
        is_bfl_control = any("query_norm.scale" in k for k in state_dict)
        if is_bfl_control:
            state_dict = _convert_bfl_flux_control_lora_to_diffusers(state_dict)
2058
2059
2060
2061
2062
2063
2064
            return cls._prepare_outputs(
                state_dict,
                metadata=metadata,
                alphas=None,
                return_alphas=return_alphas,
                return_metadata=return_lora_metadata,
            )
Aryan's avatar
Aryan committed
2065

2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
        is_fal_kontext = any("base_model" in k for k in state_dict)
        if is_fal_kontext:
            state_dict = _convert_fal_kontext_lora_to_diffusers(state_dict)
            return cls._prepare_outputs(
                state_dict,
                metadata=metadata,
                alphas=None,
                return_alphas=return_alphas,
                return_metadata=return_lora_metadata,
            )

2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
        # For state dicts like
        # https://huggingface.co/TheLastBen/Jon_Snow_Flux_LoRA
        keys = list(state_dict.keys())
        network_alphas = {}
        for k in keys:
            if "alpha" in k:
                alpha_value = state_dict.get(k)
                if (torch.is_tensor(alpha_value) and torch.is_floating_point(alpha_value)) or isinstance(
                    alpha_value, float
                ):
                    network_alphas[k] = state_dict.pop(k)
                else:
                    raise ValueError(
                        f"The alpha key ({k}) seems to be incorrect. If you think this error is unexpected, please open as issue."
                    )

2093
        if return_alphas or return_lora_metadata:
2094
2095
2096
2097
2098
2099
2100
            return cls._prepare_outputs(
                state_dict,
                metadata=metadata,
                alphas=network_alphas,
                return_alphas=return_alphas,
                return_metadata=return_lora_metadata,
            )
2101
2102
        else:
            return state_dict
Sayak Paul's avatar
Sayak Paul committed
2103
2104

    def load_lora_weights(
2105
2106
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
2107
        adapter_name: Optional[str] = None,
2108
2109
        hotswap: bool = False,
        **kwargs,
Sayak Paul's avatar
Sayak Paul committed
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`.

        All kwargs are forwarded to `self.lora_state_dict`.

        See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is
        loaded.

        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
2129
2130
2131
            low_cpu_mem_usage (`bool`, *optional*):
                `Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
2132
2133
2134
2135
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
Sayak Paul's avatar
Sayak Paul committed
2136
2137
2138
2139
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

2140
2141
2142
2143
2144
2145
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and not is_peft_version(">=", "0.13.1"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

Sayak Paul's avatar
Sayak Paul committed
2146
2147
2148
2149
2150
        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
2151
2152
        kwargs["return_lora_metadata"] = True
        state_dict, network_alphas, metadata = self.lora_state_dict(
2153
2154
            pretrained_model_name_or_path_or_dict, return_alphas=True, **kwargs
        )
Sayak Paul's avatar
Sayak Paul committed
2155

Aryan's avatar
Aryan committed
2156
2157
2158
2159
2160
2161
2162
2163
        has_lora_keys = any("lora" in key for key in state_dict.keys())

        # Flux Control LoRAs also have norm keys
        has_norm_keys = any(
            norm_key in key for key in state_dict.keys() for norm_key in self._control_lora_supported_norm_keys
        )

        if not (has_lora_keys or has_norm_keys):
Sayak Paul's avatar
Sayak Paul committed
2164
2165
            raise ValueError("Invalid LoRA checkpoint.")

Aryan's avatar
Aryan committed
2166
        transformer_lora_state_dict = {
2167
2168
2169
            k: state_dict.get(k)
            for k in list(state_dict.keys())
            if k.startswith(f"{self.transformer_name}.") and "lora" in k
Aryan's avatar
Aryan committed
2170
2171
2172
2173
        }
        transformer_norm_state_dict = {
            k: state_dict.pop(k)
            for k in list(state_dict.keys())
2174
2175
            if k.startswith(f"{self.transformer_name}.")
            and any(norm_key in k for norm_key in self._control_lora_supported_norm_keys)
Aryan's avatar
Aryan committed
2176
2177
2178
        }

        transformer = getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer
2179
2180
2181
2182
2183
        has_param_with_expanded_shape = False
        if len(transformer_lora_state_dict) > 0:
            has_param_with_expanded_shape = self._maybe_expand_transformer_param_shape_or_error_(
                transformer, transformer_lora_state_dict, transformer_norm_state_dict
            )
Aryan's avatar
Aryan committed
2184
2185
2186
2187
2188
2189
2190
2191

        if has_param_with_expanded_shape:
            logger.info(
                "The LoRA weights contain parameters that have different shapes that expected by the transformer. "
                "As a result, the state_dict of the transformer has been expanded to match the LoRA parameter shapes. "
                "To get a comprehensive list of parameter names that were modified, enable debug logging."
            )
        if len(transformer_lora_state_dict) > 0:
2192
2193
            transformer_lora_state_dict = self._maybe_expand_lora_state_dict(
                transformer=transformer, lora_state_dict=transformer_lora_state_dict
2194
            )
2195
2196
2197
2198
2199
2200
2201
2202
            for k in transformer_lora_state_dict:
                state_dict.update({k: transformer_lora_state_dict[k]})

        self.load_lora_into_transformer(
            state_dict,
            network_alphas=network_alphas,
            transformer=transformer,
            adapter_name=adapter_name,
2203
            metadata=metadata,
2204
2205
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
2206
            hotswap=hotswap,
2207
        )
Sayak Paul's avatar
Sayak Paul committed
2208

Aryan's avatar
Aryan committed
2209
2210
2211
2212
2213
2214
2215
        if len(transformer_norm_state_dict) > 0:
            transformer._transformer_norm_layers = self._load_norm_into_transformer(
                transformer_norm_state_dict,
                transformer=transformer,
                discard_original_layers=False,
            )

2216
2217
2218
2219
2220
2221
2222
        self.load_lora_into_text_encoder(
            state_dict,
            network_alphas=network_alphas,
            text_encoder=self.text_encoder,
            prefix=self.text_encoder_name,
            lora_scale=self.lora_scale,
            adapter_name=adapter_name,
2223
            metadata=metadata,
2224
2225
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
2226
            hotswap=hotswap,
2227
        )
Sayak Paul's avatar
Sayak Paul committed
2228
2229

    @classmethod
2230
    def load_lora_into_transformer(
2231
2232
2233
2234
2235
        cls,
        state_dict,
        network_alphas,
        transformer,
        adapter_name=None,
2236
        metadata=None,
2237
2238
2239
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
2240
    ):
Sayak Paul's avatar
Sayak Paul committed
2241
2242
2243
2244
2245
2246
2247
2248
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
2249
2250
2251
2252
            network_alphas (`Dict[str, float]`):
                The value of the network alpha used for stable learning and preventing underflow. This value has the
                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
2253
            transformer (`FluxTransformer2DModel`):
Sayak Paul's avatar
Sayak Paul committed
2254
2255
2256
2257
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
2258
2259
2260
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
2261
2262
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
2263
2264
2265
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
Sayak Paul's avatar
Sayak Paul committed
2266
        """
2267
2268
2269
2270
2271
        if low_cpu_mem_usage and not is_peft_version(">=", "0.13.1"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

2272
        # Load the layers corresponding to transformer.
2273
2274
2275
2276
2277
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=network_alphas,
            adapter_name=adapter_name,
2278
            metadata=metadata,
2279
2280
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
2281
            hotswap=hotswap,
2282
        )
Sayak Paul's avatar
Sayak Paul committed
2283

Aryan's avatar
Aryan committed
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
    @classmethod
    def _load_norm_into_transformer(
        cls,
        state_dict,
        transformer,
        prefix=None,
        discard_original_layers=False,
    ) -> Dict[str, torch.Tensor]:
        # Remove prefix if present
        prefix = prefix or cls.transformer_name
        for key in list(state_dict.keys()):
            if key.split(".")[0] == prefix:
2296
                state_dict[key.removeprefix(f"{prefix}.")] = state_dict.pop(key)
Aryan's avatar
Aryan committed
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337

        # Find invalid keys
        transformer_state_dict = transformer.state_dict()
        transformer_keys = set(transformer_state_dict.keys())
        state_dict_keys = set(state_dict.keys())
        extra_keys = list(state_dict_keys - transformer_keys)

        if extra_keys:
            logger.warning(
                f"Unsupported keys found in state dict when trying to load normalization layers into the transformer. The following keys will be ignored:\n{extra_keys}."
            )

        for key in extra_keys:
            state_dict.pop(key)

        # Save the layers that are going to be overwritten so that unload_lora_weights can work as expected
        overwritten_layers_state_dict = {}
        if not discard_original_layers:
            for key in state_dict.keys():
                overwritten_layers_state_dict[key] = transformer_state_dict[key].clone()

        logger.info(
            "The provided state dict contains normalization layers in addition to LoRA layers. The normalization layers will directly update the state_dict of the transformer "
            'as opposed to the LoRA layers that will co-exist separately until the "fuse_lora()" method is called. That is to say, the normalization layers will always be directly '
            "fused into the transformer and can only be unfused if `discard_original_layers=True` is passed. This might also have implications when dealing with multiple LoRAs. "
            "If you notice something unexpected, please open an issue: https://github.com/huggingface/diffusers/issues."
        )

        # We can't load with strict=True because the current state_dict does not contain all the transformer keys
        incompatible_keys = transformer.load_state_dict(state_dict, strict=False)
        unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)

        # We shouldn't expect to see the supported norm keys here being present in the unexpected keys.
        if unexpected_keys:
            if any(norm_key in k for k in unexpected_keys for norm_key in cls._control_lora_supported_norm_keys):
                raise ValueError(
                    f"Found {unexpected_keys} as unexpected keys while trying to load norm layers into the transformer."
                )

        return overwritten_layers_state_dict

Sayak Paul's avatar
Sayak Paul committed
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder
    def load_lora_into_text_encoder(
        cls,
        state_dict,
        network_alphas,
        text_encoder,
        prefix=None,
        lora_scale=1.0,
        adapter_name=None,
        _pipeline=None,
2349
        low_cpu_mem_usage=False,
2350
        hotswap: bool = False,
2351
        metadata=None,
Sayak Paul's avatar
Sayak Paul committed
2352
2353
2354
2355
2356
2357
2358
2359
2360
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `text_encoder`

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The key should be prefixed with an
                additional `text_encoder` to distinguish between unet lora layers.
            network_alphas (`Dict[str, float]`):
2361
2362
2363
                The value of the network alpha used for stable learning and preventing underflow. This value has the
                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
Sayak Paul's avatar
Sayak Paul committed
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
            text_encoder (`CLIPTextModel`):
                The text encoder model to load the LoRA layers into.
            prefix (`str`):
                Expected prefix of the `text_encoder` in the `state_dict`.
            lora_scale (`float`):
                How much to scale the output of the lora linear layer before it is added with the output of the regular
                lora layer.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
2374
2375
2376
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
2377
2378
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
2379
2380
2381
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
Sayak Paul's avatar
Sayak Paul committed
2382
        """
2383
2384
2385
2386
2387
2388
2389
2390
        _load_lora_into_text_encoder(
            state_dict=state_dict,
            network_alphas=network_alphas,
            lora_scale=lora_scale,
            text_encoder=text_encoder,
            prefix=prefix,
            text_encoder_name=cls.text_encoder_name,
            adapter_name=adapter_name,
2391
            metadata=metadata,
2392
2393
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
2394
            hotswap=hotswap,
2395
        )
Sayak Paul's avatar
Sayak Paul committed
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.save_lora_weights with unet->transformer
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        text_encoder_lora_layers: Dict[str, torch.nn.Module] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
2408
2409
        transformer_lora_adapter_metadata=None,
        text_encoder_lora_adapter_metadata=None,
Sayak Paul's avatar
Sayak Paul committed
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
    ):
        r"""
        Save the LoRA parameters corresponding to the UNet and text encoder.

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
                encoder LoRA state dict because it comes from 🤗 Transformers.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
2432
2433
2434
2435
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
            text_encoder_lora_adapter_metadata:
                LoRA adapter metadata associated with the text encoder to be serialized with the state dict.
Sayak Paul's avatar
Sayak Paul committed
2436
2437
        """
        state_dict = {}
2438
        lora_adapter_metadata = {}
Sayak Paul's avatar
Sayak Paul committed
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448

        if not (transformer_lora_layers or text_encoder_lora_layers):
            raise ValueError("You must pass at least one of `transformer_lora_layers` and `text_encoder_lora_layers`.")

        if transformer_lora_layers:
            state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if text_encoder_lora_layers:
            state_dict.update(cls.pack_weights(text_encoder_lora_layers, cls.text_encoder_name))

2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
        if transformer_lora_adapter_metadata:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )

        if text_encoder_lora_adapter_metadata:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(text_encoder_lora_adapter_metadata, cls.text_encoder_name)
            )

Sayak Paul's avatar
Sayak Paul committed
2459
2460
2461
2462
2463
2464
2465
2466
        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
2467
            lora_adapter_metadata=lora_adapter_metadata,
Sayak Paul's avatar
Sayak Paul committed
2468
2469
2470
2471
        )

    def fuse_lora(
        self,
2472
        components: List[str] = ["transformer"],
Sayak Paul's avatar
Sayak Paul committed
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
Aryan's avatar
Aryan committed
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521

        transformer = getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer
        if (
            hasattr(transformer, "_transformer_norm_layers")
            and isinstance(transformer._transformer_norm_layers, dict)
            and len(transformer._transformer_norm_layers.keys()) > 0
        ):
            logger.info(
                "The provided state dict contains normalization layers in addition to LoRA layers. The normalization layers will be directly updated the state_dict of the transformer "
                "as opposed to the LoRA layers that will co-exist separately until the 'fuse_lora()' method is called. That is to say, the normalization layers will always be directly "
                "fused into the transformer and can only be unfused if `discard_original_layers=True` is passed."
            )

Sayak Paul's avatar
Sayak Paul committed
2522
        super().fuse_lora(
2523
2524
2525
2526
2527
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
Sayak Paul's avatar
Sayak Paul committed
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
        )

    def unfuse_lora(self, components: List[str] = ["transformer", "text_encoder"], **kwargs):
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
        """
Aryan's avatar
Aryan committed
2544
2545
2546
2547
        transformer = getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer
        if hasattr(transformer, "_transformer_norm_layers") and transformer._transformer_norm_layers:
            transformer.load_state_dict(transformer._transformer_norm_layers, strict=False)

2548
        super().unfuse_lora(components=components, **kwargs)
Sayak Paul's avatar
Sayak Paul committed
2549

2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
    # We override this here account for `_transformer_norm_layers` and `_overwritten_params`.
    def unload_lora_weights(self, reset_to_overwritten_params=False):
        """
        Unloads the LoRA parameters.

        Args:
            reset_to_overwritten_params (`bool`, defaults to `False`): Whether to reset the LoRA-loaded modules
                to their original params. Refer to the [Flux
                documentation](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux) to learn more.

        Examples:

        ```python
        >>> # Assuming `pipeline` is already loaded with the LoRA parameters.
        >>> pipeline.unload_lora_weights()
        >>> ...
        ```
        """
Aryan's avatar
Aryan committed
2568
2569
2570
2571
2572
2573
2574
        super().unload_lora_weights()

        transformer = getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer
        if hasattr(transformer, "_transformer_norm_layers") and transformer._transformer_norm_layers:
            transformer.load_state_dict(transformer._transformer_norm_layers, strict=False)
            transformer._transformer_norm_layers = None

2575
        if reset_to_overwritten_params and getattr(transformer, "_overwritten_params", None) is not None:
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
            overwritten_params = transformer._overwritten_params
            module_names = set()

            for param_name in overwritten_params:
                if param_name.endswith(".weight"):
                    module_names.add(param_name.replace(".weight", ""))

            for name, module in transformer.named_modules():
                if isinstance(module, torch.nn.Linear) and name in module_names:
                    module_weight = module.weight.data
                    module_bias = module.bias.data if module.bias is not None else None
                    bias = module_bias is not None

                    parent_module_name, _, current_module_name = name.rpartition(".")
                    parent_module = transformer.get_submodule(parent_module_name)

                    current_param_weight = overwritten_params[f"{name}.weight"]
                    in_features, out_features = current_param_weight.shape[1], current_param_weight.shape[0]
                    with torch.device("meta"):
                        original_module = torch.nn.Linear(
                            in_features,
                            out_features,
                            bias=bias,
                            dtype=module_weight.dtype,
                        )

                    tmp_state_dict = {"weight": current_param_weight}
                    if module_bias is not None:
                        tmp_state_dict.update({"bias": overwritten_params[f"{name}.bias"]})
                    original_module.load_state_dict(tmp_state_dict, assign=True, strict=True)
                    setattr(parent_module, current_module_name, original_module)

                    del tmp_state_dict

                    if current_module_name in _MODULE_NAME_TO_ATTRIBUTE_MAP_FLUX:
                        attribute_name = _MODULE_NAME_TO_ATTRIBUTE_MAP_FLUX[current_module_name]
                        new_value = int(current_param_weight.shape[1])
                        old_value = getattr(transformer.config, attribute_name)
                        setattr(transformer.config, attribute_name, new_value)
                        logger.info(
                            f"Set the {attribute_name} attribute of the model to {new_value} from {old_value}."
                        )

Aryan's avatar
Aryan committed
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
    @classmethod
    def _maybe_expand_transformer_param_shape_or_error_(
        cls,
        transformer: torch.nn.Module,
        lora_state_dict=None,
        norm_state_dict=None,
        prefix=None,
    ) -> bool:
        """
        Control LoRA expands the shape of the input layer from (3072, 64) to (3072, 128). This method handles that and
2629
        generalizes things a bit so that any parameter that needs expansion receives appropriate treatment.
Aryan's avatar
Aryan committed
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
        """
        state_dict = {}
        if lora_state_dict is not None:
            state_dict.update(lora_state_dict)
        if norm_state_dict is not None:
            state_dict.update(norm_state_dict)

        # Remove prefix if present
        prefix = prefix or cls.transformer_name
        for key in list(state_dict.keys()):
            if key.split(".")[0] == prefix:
2641
                state_dict[key.removeprefix(f"{prefix}.")] = state_dict.pop(key)
Aryan's avatar
Aryan committed
2642
2643
2644

        # Expand transformer parameter shapes if they don't match lora
        has_param_with_shape_update = False
2645
2646
        overwritten_params = {}

2647
        is_peft_loaded = getattr(transformer, "peft_config", None) is not None
hlky's avatar
hlky committed
2648
        is_quantized = hasattr(transformer, "hf_quantizer")
Aryan's avatar
Aryan committed
2649
2650
2651
        for name, module in transformer.named_modules():
            if isinstance(module, torch.nn.Linear):
                module_weight = module.weight.data
2652
                module_bias = module.bias.data if module.bias is not None else None
Aryan's avatar
Aryan committed
2653
2654
                bias = module_bias is not None

2655
2656
2657
2658
                lora_base_name = name.replace(".base_layer", "") if is_peft_loaded else name
                lora_A_weight_name = f"{lora_base_name}.lora_A.weight"
                lora_B_weight_name = f"{lora_base_name}.lora_B.weight"
                if lora_A_weight_name not in state_dict:
Aryan's avatar
Aryan committed
2659
2660
2661
2662
2663
                    continue

                in_features = state_dict[lora_A_weight_name].shape[1]
                out_features = state_dict[lora_B_weight_name].shape[0]

2664
2665
2666
2667
2668
                # Model maybe loaded with different quantization schemes which may flatten the params.
                # `bitsandbytes`, for example, flatten the weights when using 4bit. 8bit bnb models
                # preserve weight shape.
                module_weight_shape = cls._calculate_module_shape(model=transformer, base_module=module)

Aryan's avatar
Aryan committed
2669
                # This means there's no need for an expansion in the params, so we simply skip.
2670
                if tuple(module_weight_shape) == (out_features, in_features):
Aryan's avatar
Aryan committed
2671
2672
                    continue

hlky's avatar
hlky committed
2673
                module_out_features, module_in_features = module_weight_shape
2674
2675
2676
2677
2678
2679
                debug_message = ""
                if in_features > module_in_features:
                    debug_message += (
                        f'Expanding the nn.Linear input/output features for module="{name}" because the provided LoRA '
                        f"checkpoint contains higher number of features than expected. The number of input_features will be "
                        f"expanded from {module_in_features} to {in_features}"
Aryan's avatar
Aryan committed
2680
                    )
2681
                if out_features > module_out_features:
2682
2683
2684
2685
2686
2687
                    debug_message += (
                        ", and the number of output features will be "
                        f"expanded from {module_out_features} to {out_features}."
                    )
                else:
                    debug_message += "."
2688
2689
2690
2691
2692
2693
2694
2695
                if debug_message:
                    logger.debug(debug_message)

                if out_features > module_out_features or in_features > module_in_features:
                    has_param_with_shape_update = True
                    parent_module_name, _, current_module_name = name.rpartition(".")
                    parent_module = transformer.get_submodule(parent_module_name)

hlky's avatar
hlky committed
2696
2697
2698
2699
                    if is_quantized:
                        module_weight = _maybe_dequantize_weight_for_expanded_lora(transformer, module)

                    # TODO: consider if this layer needs to be a quantized layer as well if `is_quantized` is True.
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
                    with torch.device("meta"):
                        expanded_module = torch.nn.Linear(
                            in_features, out_features, bias=bias, dtype=module_weight.dtype
                        )
                    # Only weights are expanded and biases are not. This is because only the input dimensions
                    # are changed while the output dimensions remain the same. The shape of the weight tensor
                    # is (out_features, in_features), while the shape of bias tensor is (out_features,), which
                    # explains the reason why only weights are expanded.
                    new_weight = torch.zeros_like(
                        expanded_module.weight.data, device=module_weight.device, dtype=module_weight.dtype
                    )
hlky's avatar
hlky committed
2711
                    slices = tuple(slice(0, dim) for dim in module_weight_shape)
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
                    new_weight[slices] = module_weight
                    tmp_state_dict = {"weight": new_weight}
                    if module_bias is not None:
                        tmp_state_dict["bias"] = module_bias
                    expanded_module.load_state_dict(tmp_state_dict, strict=True, assign=True)

                    setattr(parent_module, current_module_name, expanded_module)

                    del tmp_state_dict

                    if current_module_name in _MODULE_NAME_TO_ATTRIBUTE_MAP_FLUX:
                        attribute_name = _MODULE_NAME_TO_ATTRIBUTE_MAP_FLUX[current_module_name]
                        new_value = int(expanded_module.weight.data.shape[1])
                        old_value = getattr(transformer.config, attribute_name)
                        setattr(transformer.config, attribute_name, new_value)
                        logger.info(
                            f"Set the {attribute_name} attribute of the model to {new_value} from {old_value}."
                        )
Aryan's avatar
Aryan committed
2730

2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
                    # For `unload_lora_weights()`.
                    # TODO: this could lead to more memory overhead if the number of overwritten params
                    # are large. Should be revisited later and tackled through a `discard_original_layers` arg.
                    overwritten_params[f"{current_module_name}.weight"] = module_weight
                    if module_bias is not None:
                        overwritten_params[f"{current_module_name}.bias"] = module_bias

        if len(overwritten_params) > 0:
            transformer._overwritten_params = overwritten_params

2741
        return has_param_with_shape_update
Aryan's avatar
Aryan committed
2742

2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
    @classmethod
    def _maybe_expand_lora_state_dict(cls, transformer, lora_state_dict):
        expanded_module_names = set()
        transformer_state_dict = transformer.state_dict()
        prefix = f"{cls.transformer_name}."

        lora_module_names = [
            key[: -len(".lora_A.weight")] for key in lora_state_dict if key.endswith(".lora_A.weight")
        ]
        lora_module_names = [name[len(prefix) :] for name in lora_module_names if name.startswith(prefix)]
        lora_module_names = sorted(set(lora_module_names))
        transformer_module_names = sorted({name for name, _ in transformer.named_modules()})
        unexpected_modules = set(lora_module_names) - set(transformer_module_names)
        if unexpected_modules:
            logger.debug(f"Found unexpected modules: {unexpected_modules}. These will be ignored.")

        for k in lora_module_names:
            if k in unexpected_modules:
                continue

            base_param_name = (
2764
                f"{k.replace(prefix, '')}.base_layer.weight"
2765
                if f"{k.replace(prefix, '')}.base_layer.weight" in transformer_state_dict
2766
                else f"{k.replace(prefix, '')}.weight"
2767
2768
2769
2770
            )
            base_weight_param = transformer_state_dict[base_param_name]
            lora_A_param = lora_state_dict[f"{prefix}{k}.lora_A.weight"]

2771
2772
2773
2774
            # TODO (sayakpaul): Handle the cases when we actually need to expand when using quantization.
            base_module_shape = cls._calculate_module_shape(model=transformer, base_weight_param_name=base_param_name)

            if base_module_shape[1] > lora_A_param.shape[1]:
2775
2776
2777
2778
2779
                shape = (lora_A_param.shape[0], base_weight_param.shape[1])
                expanded_state_dict_weight = torch.zeros(shape, device=base_weight_param.device)
                expanded_state_dict_weight[:, : lora_A_param.shape[1]].copy_(lora_A_param)
                lora_state_dict[f"{prefix}{k}.lora_A.weight"] = expanded_state_dict_weight
                expanded_module_names.add(k)
2780
            elif base_module_shape[1] < lora_A_param.shape[1]:
2781
2782
                raise NotImplementedError(
                    f"This LoRA param ({k}.lora_A.weight) has an incompatible shape {lora_A_param.shape}. Please open an issue to file for a feature request - https://github.com/huggingface/diffusers/issues/new."
Aryan's avatar
Aryan committed
2783
2784
                )

2785
2786
2787
2788
        if expanded_module_names:
            logger.info(
                f"The following LoRA modules were zero padded to match the state dict of {cls.transformer_name}: {expanded_module_names}. Please open an issue if you think this was unexpected - https://github.com/huggingface/diffusers/issues/new."
            )
Aryan's avatar
Aryan committed
2789

2790
        return lora_state_dict
Aryan's avatar
Aryan committed
2791

2792
2793
2794
2795
2796
2797
2798
    @staticmethod
    def _calculate_module_shape(
        model: "torch.nn.Module",
        base_module: "torch.nn.Linear" = None,
        base_weight_param_name: str = None,
    ) -> "torch.Size":
        def _get_weight_shape(weight: torch.Tensor):
hlky's avatar
hlky committed
2799
2800
2801
2802
2803
2804
            if weight.__class__.__name__ == "Params4bit":
                return weight.quant_state.shape
            elif weight.__class__.__name__ == "GGUFParameter":
                return weight.quant_shape
            else:
                return weight.shape
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818

        if base_module is not None:
            return _get_weight_shape(base_module.weight)
        elif base_weight_param_name is not None:
            if not base_weight_param_name.endswith(".weight"):
                raise ValueError(
                    f"Invalid `base_weight_param_name` passed as it does not end with '.weight' {base_weight_param_name=}."
                )
            module_path = base_weight_param_name.rsplit(".weight", 1)[0]
            submodule = get_submodule_by_name(model, module_path)
            return _get_weight_shape(submodule.weight)

        raise ValueError("Either `base_module` or `base_weight_param_name` must be provided.")

2819
2820
2821
2822
2823
2824
2825
2826
2827
    @staticmethod
    def _prepare_outputs(state_dict, metadata, alphas=None, return_alphas=False, return_metadata=False):
        outputs = [state_dict]
        if return_alphas:
            outputs.append(alphas)
        if return_metadata:
            outputs.append(metadata)
        return tuple(outputs) if (return_alphas or return_metadata) else state_dict

Sayak Paul's avatar
Sayak Paul committed
2828

2829
2830
2831
2832
2833
2834
# The reason why we subclass from `StableDiffusionLoraLoaderMixin` here is because Amused initially
# relied on `StableDiffusionLoraLoaderMixin` for its LoRA support.
class AmusedLoraLoaderMixin(StableDiffusionLoraLoaderMixin):
    _lora_loadable_modules = ["transformer", "text_encoder"]
    transformer_name = TRANSFORMER_NAME
    text_encoder_name = TEXT_ENCODER_NAME
Dhruv Nair's avatar
Dhruv Nair committed
2835
2836

    @classmethod
2837
2838
    # Copied from diffusers.loaders.lora_pipeline.FluxLoraLoaderMixin.load_lora_into_transformer with FluxTransformer2DModel->UVit2DModel
    def load_lora_into_transformer(
2839
2840
2841
2842
2843
        cls,
        state_dict,
        network_alphas,
        transformer,
        adapter_name=None,
2844
        metadata=None,
2845
2846
2847
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
2848
    ):
Dhruv Nair's avatar
Dhruv Nair committed
2849
2850
2851
2852
2853
2854
2855
2856
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
2857
            network_alphas (`Dict[str, float]`):
2858
                The value of the network alpha used for stable learning and preventing underflow. This value has the
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
            transformer (`UVit2DModel`):
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
2871
2872
2873
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
Dhruv Nair's avatar
Dhruv Nair committed
2874
        """
2875
2876
2877
2878
        if low_cpu_mem_usage and not is_peft_version(">=", "0.13.1"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )
Dhruv Nair's avatar
Dhruv Nair committed
2879

2880
        # Load the layers corresponding to transformer.
2881
2882
2883
2884
2885
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=network_alphas,
            adapter_name=adapter_name,
2886
            metadata=metadata,
2887
2888
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
2889
            hotswap=hotswap,
2890
        )
Dhruv Nair's avatar
Dhruv Nair committed
2891

2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder
    def load_lora_into_text_encoder(
        cls,
        state_dict,
        network_alphas,
        text_encoder,
        prefix=None,
        lora_scale=1.0,
        adapter_name=None,
        _pipeline=None,
2903
        low_cpu_mem_usage=False,
2904
        hotswap: bool = False,
2905
        metadata=None,
2906
2907
2908
2909
2910
2911
2912
2913
2914
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `text_encoder`

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The key should be prefixed with an
                additional `text_encoder` to distinguish between unet lora layers.
            network_alphas (`Dict[str, float]`):
2915
2916
2917
                The value of the network alpha used for stable learning and preventing underflow. This value has the
                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
            text_encoder (`CLIPTextModel`):
                The text encoder model to load the LoRA layers into.
            prefix (`str`):
                Expected prefix of the `text_encoder` in the `state_dict`.
            lora_scale (`float`):
                How much to scale the output of the lora linear layer before it is added with the output of the regular
                lora layer.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
2928
2929
2930
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
2931
2932
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
2933
2934
2935
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
2936
        """
2937
2938
2939
2940
2941
2942
2943
2944
        _load_lora_into_text_encoder(
            state_dict=state_dict,
            network_alphas=network_alphas,
            lora_scale=lora_scale,
            text_encoder=text_encoder,
            prefix=prefix,
            text_encoder_name=cls.text_encoder_name,
            adapter_name=adapter_name,
2945
            metadata=metadata,
2946
2947
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
2948
            hotswap=hotswap,
2949
        )
2950

Dhruv Nair's avatar
Dhruv Nair committed
2951
2952
2953
2954
    @classmethod
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
2955
        text_encoder_lora_layers: Dict[str, torch.nn.Module] = None,
Dhruv Nair's avatar
Dhruv Nair committed
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
        transformer_lora_layers: Dict[str, torch.nn.Module] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
    ):
        r"""
        Save the LoRA parameters corresponding to the UNet and text encoder.

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
2968
2969
2970
2971
2972
            unet_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `unet`.
            text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
                encoder LoRA state dict because it comes from 🤗 Transformers.
Dhruv Nair's avatar
Dhruv Nair committed
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
        """
        state_dict = {}

2986
2987
        if not (transformer_lora_layers or text_encoder_lora_layers):
            raise ValueError("You must pass at least one of `transformer_lora_layers` or `text_encoder_lora_layers`.")
Dhruv Nair's avatar
Dhruv Nair committed
2988
2989

        if transformer_lora_layers:
2990
            state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))
Dhruv Nair's avatar
Dhruv Nair committed
2991

2992
        if text_encoder_lora_layers:
2993
            state_dict.update(cls.pack_weights(text_encoder_lora_layers, cls.text_encoder_name))
2994

Dhruv Nair's avatar
Dhruv Nair committed
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
        )

3005

Aryan's avatar
Aryan committed
3006
3007
class CogVideoXLoraLoaderMixin(LoraBaseMixin):
    r"""
3008
    Load LoRA layers into [`CogVideoXTransformer3DModel`]. Specific to [`CogVideoXPipeline`].
Aryan's avatar
Aryan committed
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.lora_state_dict
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
3065
3066
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
Aryan's avatar
Aryan committed
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079

        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
3080
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
Aryan's avatar
Aryan committed
3081
3082
3083
3084
3085
3086

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

3087
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
Aryan's avatar
Aryan committed
3088

3089
        state_dict, metadata = _fetch_state_dict(
Aryan's avatar
Aryan committed
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

Sayak Paul's avatar
Sayak Paul committed
3104
3105
3106
3107
3108
3109
        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

3110
3111
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
Aryan's avatar
Aryan committed
3112
3113

    def load_lora_weights(
3114
3115
3116
3117
3118
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
Aryan's avatar
Aryan committed
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
3133
3134
3135
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
3136
3137
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
3138
3139
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
Aryan's avatar
Aryan committed
3140
3141
3142
3143
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

3144
3145
3146
3147
3148
3149
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

Aryan's avatar
Aryan committed
3150
3151
3152
3153
3154
        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
3155
3156
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
Aryan's avatar
Aryan committed
3157

Sayak Paul's avatar
Sayak Paul committed
3158
        is_correct_format = all("lora" in key for key in state_dict.keys())
Aryan's avatar
Aryan committed
3159
3160
3161
3162
3163
3164
3165
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
3166
            metadata=metadata,
Aryan's avatar
Aryan committed
3167
            _pipeline=self,
3168
            low_cpu_mem_usage=low_cpu_mem_usage,
3169
            hotswap=hotswap,
Aryan's avatar
Aryan committed
3170
3171
3172
        )

    @classmethod
3173
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->CogVideoXTransformer3DModel
3174
    def load_lora_into_transformer(
3175
3176
3177
3178
3179
3180
3181
3182
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
3183
    ):
Aryan's avatar
Aryan committed
3184
3185
3186
3187
3188
3189
3190
3191
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
3192
            transformer (`CogVideoXTransformer3DModel`):
Aryan's avatar
Aryan committed
3193
3194
3195
3196
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
3197
3198
3199
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
3200
3201
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
3202
3203
3204
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
Aryan's avatar
Aryan committed
3205
        """
3206
3207
3208
3209
3210
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

3211
3212
3213
3214
3215
3216
        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
3217
            metadata=metadata,
3218
3219
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
3220
            hotswap=hotswap,
3221
        )
Aryan's avatar
Aryan committed
3222
3223
3224
3225
3226
3227
3228

    @classmethod
    # Adapted from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.save_lora_weights without support for text encoder
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
3229
3230
3231
3232
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
3233
        transformer_lora_adapter_metadata: Optional[dict] = None,
3234
3235
    ):
        r"""
3236
        Save the LoRA parameters corresponding to the transformer.
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
3253
3254
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
3255
3256
        """
        state_dict = {}
3257
        lora_adapter_metadata = {}
3258
3259
3260
3261

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

3262
3263
3264
3265
3266
3267
        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )
3268
3269
3270
3271
3272
3273
3274
3275
3276

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
3277
            lora_adapter_metadata=lora_adapter_metadata,
3278
3279
3280
3281
        )

    def fuse_lora(
        self,
3282
        components: List[str] = ["transformer"],
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
3320
3321
3322
3323
3324
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
3325
3326
        )

3327
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
3342
        super().unfuse_lora(components=components, **kwargs)
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403


class Mochi1LoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`MochiTransformer3DModel`]. Specific to [`MochiPipeline`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.lora_state_dict
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
3404
3405
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418

        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
3419
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
3420
3421
3422
3423
3424
3425

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

3426
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
3427

3428
        state_dict, metadata = _fetch_state_dict(
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

3449
3450
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
3451
3452
3453

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
    def load_lora_weights(
3454
3455
3456
3457
3458
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
3476
3477
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
3495
3496
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
3497
3498
3499
3500
3501
3502
3503
3504
3505

        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
3506
            metadata=metadata,
3507
3508
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
3509
            hotswap=hotswap,
3510
3511
3512
        )

    @classmethod
3513
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->MochiTransformer3DModel
3514
    def load_lora_into_transformer(
3515
3516
3517
3518
3519
3520
3521
3522
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
3523
3524
3525
3526
3527
3528
3529
3530
3531
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
3532
            transformer (`MochiTransformer3DModel`):
3533
3534
3535
3536
3537
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
Aryan's avatar
Aryan committed
3538
3539
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
3540
3541
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
3542
3543
3544
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
Aryan's avatar
Aryan committed
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
3557
            metadata=metadata,
Aryan's avatar
Aryan committed
3558
3559
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
3560
            hotswap=hotswap,
Aryan's avatar
Aryan committed
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
3573
        transformer_lora_adapter_metadata: Optional[dict] = None,
Aryan's avatar
Aryan committed
3574
3575
    ):
        r"""
3576
        Save the LoRA parameters corresponding to the transformer.
Aryan's avatar
Aryan committed
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
3593
3594
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
Aryan's avatar
Aryan committed
3595
3596
        """
        state_dict = {}
3597
        lora_adapter_metadata = {}
Aryan's avatar
Aryan committed
3598
3599
3600
3601

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

3602
3603
3604
3605
3606
3607
        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )
Aryan's avatar
Aryan committed
3608
3609
3610
3611
3612
3613
3614
3615
3616

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
3617
            lora_adapter_metadata=lora_adapter_metadata,
Aryan's avatar
Aryan committed
3618
3619
        )

3620
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
Aryan's avatar
Aryan committed
3621
3622
    def fuse_lora(
        self,
3623
        components: List[str] = ["transformer"],
Aryan's avatar
Aryan committed
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
3661
3662
3663
3664
3665
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
Aryan's avatar
Aryan committed
3666
3667
        )

3668
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
3669
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
Aryan's avatar
Aryan committed
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
3684
        super().unfuse_lora(components=components, **kwargs)
Aryan's avatar
Aryan committed
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744


class LTXVideoLoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`LTXVideoTransformer3DModel`]. Specific to [`LTXPipeline`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
3745
3746
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
Aryan's avatar
Aryan committed
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
3759
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
Aryan's avatar
Aryan committed
3760
3761
3762
3763
3764
3765

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

3766
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
Aryan's avatar
Aryan committed
3767

3768
        state_dict, metadata = _fetch_state_dict(
Aryan's avatar
Aryan committed
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

3789
3790
3791
3792
        is_non_diffusers_format = any(k.startswith("diffusion_model.") for k in state_dict)
        if is_non_diffusers_format:
            state_dict = _convert_non_diffusers_ltxv_lora_to_diffusers(state_dict)

3793
3794
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
Aryan's avatar
Aryan committed
3795
3796
3797

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
    def load_lora_weights(
3798
3799
3800
3801
3802
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
Aryan's avatar
Aryan committed
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
3820
3821
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
Aryan's avatar
Aryan committed
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
3839
3840
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
Aryan's avatar
Aryan committed
3841
3842
3843
3844
3845
3846
3847
3848
3849

        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
3850
            metadata=metadata,
Aryan's avatar
Aryan committed
3851
3852
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
3853
            hotswap=hotswap,
Aryan's avatar
Aryan committed
3854
3855
3856
3857
3858
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->LTXVideoTransformer3DModel
    def load_lora_into_transformer(
3859
3860
3861
3862
3863
3864
3865
3866
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
Aryan's avatar
Aryan committed
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`LTXVideoTransformer3DModel`):
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
3882
3883
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
3884
3885
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
3886
3887
3888
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
3901
            metadata=metadata,
3902
3903
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
3904
            hotswap=hotswap,
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
3917
        transformer_lora_adapter_metadata: Optional[dict] = None,
3918
3919
    ):
        r"""
3920
        Save the LoRA parameters corresponding to the transformer.
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
3937
3938
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
3939
3940
        """
        state_dict = {}
3941
        lora_adapter_metadata = {}
3942
3943
3944
3945

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

3946
3947
3948
3949
3950
3951
        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )
3952
3953
3954
3955
3956
3957
3958
3959
3960

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
3961
            lora_adapter_metadata=lora_adapter_metadata,
3962
3963
        )

3964
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
3965
3966
    def fuse_lora(
        self,
3967
        components: List[str] = ["transformer"],
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
4005
4006
4007
4008
4009
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
4010
4011
        )

4012
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
4013
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
4028
        super().unfuse_lora(components=components, **kwargs)
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089


class SanaLoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`SanaTransformer2DModel`]. Specific to [`SanaPipeline`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.lora_state_dict
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
4090
4091
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104

        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
4105
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
4106
4107
4108
4109
4110
4111

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

4112
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
4113

4114
        state_dict, metadata = _fetch_state_dict(
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

4135
4136
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
4137
4138
4139

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
    def load_lora_weights(
4140
4141
4142
4143
4144
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
4162
4163
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
4181
4182
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
4183
4184
4185
4186
4187
4188
4189
4190
4191

        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
4192
            metadata=metadata,
4193
4194
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
4195
            hotswap=hotswap,
4196
4197
4198
4199
4200
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->SanaTransformer2DModel
    def load_lora_into_transformer(
4201
4202
4203
4204
4205
4206
4207
4208
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`SanaTransformer2DModel`):
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
4224
4225
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
4226
4227
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
4228
4229
4230
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
4243
            metadata=metadata,
4244
4245
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
4246
            hotswap=hotswap,
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
4259
        transformer_lora_adapter_metadata: Optional[dict] = None,
4260
4261
    ):
        r"""
4262
        Save the LoRA parameters corresponding to the transformer.
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
4279
4280
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
4281
4282
        """
        state_dict = {}
4283
        lora_adapter_metadata = {}
4284
4285
4286
4287

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

4288
4289
4290
4291
4292
4293
        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )
4294
4295
4296
4297
4298
4299
4300
4301
4302

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
4303
            lora_adapter_metadata=lora_adapter_metadata,
4304
4305
        )

4306
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
4307
4308
    def fuse_lora(
        self,
4309
        components: List[str] = ["transformer"],
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
4347
4348
4349
4350
4351
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
4352
4353
        )

4354
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
4355
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
4370
        super().unfuse_lora(components=components, **kwargs)
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392


class HunyuanVideoLoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`HunyuanVideoTransformer3DModel`]. Specific to [`HunyuanVideoPipeline`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

4393
        We support loading original format HunyuanVideo LoRA checkpoints.
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
4431
4432
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
4445
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
4446
4447
4448
4449
4450
4451

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

4452
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
4453

4454
        state_dict, metadata = _fetch_state_dict(
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

4475
4476
4477
4478
        is_original_hunyuan_video = any("img_attn_qkv" in k for k in state_dict)
        if is_original_hunyuan_video:
            state_dict = _convert_hunyuan_video_lora_to_diffusers(state_dict)

4479
4480
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
4481
4482
4483

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
    def load_lora_weights(
4484
4485
4486
4487
4488
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
4506
4507
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
4525
4526
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
4527
4528
4529
4530
4531
4532
4533
4534
4535

        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
4536
            metadata=metadata,
4537
4538
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
4539
            hotswap=hotswap,
4540
4541
4542
4543
4544
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->HunyuanVideoTransformer3DModel
    def load_lora_into_transformer(
4545
4546
4547
4548
4549
4550
4551
4552
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`HunyuanVideoTransformer3DModel`):
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
4568
4569
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
4570
4571
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
4572
4573
4574
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
4587
            metadata=metadata,
4588
4589
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
4590
            hotswap=hotswap,
4591
4592
4593
4594
4595
4596
4597
4598
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
Aryan's avatar
Aryan committed
4599
4600
4601
4602
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
4603
        transformer_lora_adapter_metadata: Optional[dict] = None,
Aryan's avatar
Aryan committed
4604
4605
    ):
        r"""
4606
        Save the LoRA parameters corresponding to the transformer.
Aryan's avatar
Aryan committed
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
4623
4624
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
Aryan's avatar
Aryan committed
4625
4626
        """
        state_dict = {}
4627
        lora_adapter_metadata = {}
Aryan's avatar
Aryan committed
4628
4629
4630
4631

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

4632
4633
4634
4635
4636
4637
        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )
Aryan's avatar
Aryan committed
4638
4639
4640
4641
4642
4643
4644
4645
4646

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
4647
            lora_adapter_metadata=lora_adapter_metadata,
Aryan's avatar
Aryan committed
4648
4649
        )

4650
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
Aryan's avatar
Aryan committed
4651
4652
    def fuse_lora(
        self,
4653
        components: List[str] = ["transformer"],
Aryan's avatar
Aryan committed
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
4690
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
4691
4692
4693
4694
4695
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
Aryan's avatar
Aryan committed
4696
4697
        )

4698
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
4699
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
Aryan's avatar
Aryan committed
4700
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
4714
        super().unfuse_lora(components=components, **kwargs)
Aryan's avatar
Aryan committed
4715
4716


4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
class Lumina2LoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`Lumina2Transformer2DModel`]. Specific to [`Lumina2Text2ImgPipeline`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
4775
4776
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
4789
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
4790
4791
4792
4793
4794
4795

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

4796
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
4797

4798
        state_dict, metadata = _fetch_state_dict(
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

4819
4820
4821
4822
4823
        # conversion.
        non_diffusers = any(k.startswith("diffusion_model.") for k in state_dict)
        if non_diffusers:
            state_dict = _convert_non_diffusers_lumina2_lora_to_diffusers(state_dict)

4824
4825
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
4826
4827
4828

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
    def load_lora_weights(
4829
4830
4831
4832
4833
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847
4848
4849
4850
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
4851
4852
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
4870
4871
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
4872
4873
4874
4875
4876
4877
4878
4879
4880

        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
4881
            metadata=metadata,
4882
4883
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
4884
            hotswap=hotswap,
4885
4886
4887
4888
4889
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->Lumina2Transformer2DModel
    def load_lora_into_transformer(
4890
4891
4892
4893
4894
4895
4896
4897
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
4898
4899
4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`Lumina2Transformer2DModel`):
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
4915
4916
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
4917
4918
4919
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
4920
4921
4922
4923
4924
4925
4926
4927
4928
4929
4930
4931
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
4932
            metadata=metadata,
4933
4934
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
4935
            hotswap=hotswap,
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
4948
        transformer_lora_adapter_metadata: Optional[dict] = None,
4949
4950
    ):
        r"""
4951
        Save the LoRA parameters corresponding to the transformer.
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
4968
4969
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
4970
4971
        """
        state_dict = {}
4972
        lora_adapter_metadata = {}
4973
4974
4975
4976

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

4977
4978
4979
4980
4981
4982
        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )
4983
4984
4985
4986
4987
4988
4989
4990
4991

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
4992
            lora_adapter_metadata=lora_adapter_metadata,
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
5028
5029
5030
5031
5032
5033
5034
5035
        )

    # Copied from diffusers.loaders.lora_pipeline.SanaLoraLoaderMixin.fuse_lora
    def fuse_lora(
        self,
        components: List[str] = ["transformer"],
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
5036
5037
5038
5039
5040
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
5041
5042
5043
5044
5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
5057
5058
        )

    # Copied from diffusers.loaders.lora_pipeline.SanaLoraLoaderMixin.unfuse_lora
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
5059
        super().unfuse_lora(components=components, **kwargs)
5060
5061


Aryan's avatar
Aryan committed
5062
5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
5074
5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
5085
5086
5087
5088
5089
5090
5091
5092
5093
5094
5095
5096
5097
5098
5099
5100
5101
5102
5103
5104
5105
5106
5107
5108
5109
5110
5111
5112
5113
5114
5115
5116
5117
5118
5119
class WanLoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`WanTransformer3DModel`]. Specific to [`WanPipeline`] and `[WanImageToVideoPipeline`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
5120
5121
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
Aryan's avatar
Aryan committed
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
5134
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
Aryan's avatar
Aryan committed
5135
5136
5137
5138
5139
5140

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

5141
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
Aryan's avatar
Aryan committed
5142

5143
        state_dict, metadata = _fetch_state_dict(
Aryan's avatar
Aryan committed
5144
5145
5146
5147
5148
5149
5150
5151
5152
5153
5154
5155
5156
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )
5157
5158
        if any(k.startswith("diffusion_model.") for k in state_dict):
            state_dict = _convert_non_diffusers_wan_lora_to_diffusers(state_dict)
5159
5160
        elif any(k.startswith("lora_unet_") for k in state_dict):
            state_dict = _convert_musubi_wan_lora_to_diffusers(state_dict)
Aryan's avatar
Aryan committed
5161
5162
5163
5164
5165
5166
5167

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

5168
5169
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
Aryan's avatar
Aryan committed
5170

5171
5172
5173
5174
5175
5176
5177
5178
5179
    @classmethod
    def _maybe_expand_t2v_lora_for_i2v(
        cls,
        transformer: torch.nn.Module,
        state_dict,
    ):
        if transformer.config.image_dim is None:
            return state_dict

5180
5181
        target_device = transformer.device

5182
        if any(k.startswith("transformer.blocks.") for k in state_dict):
5183
            num_blocks = len({k.split("blocks.")[1].split(".")[0] for k in state_dict if "blocks." in k})
5184
            is_i2v_lora = any("add_k_proj" in k for k in state_dict) and any("add_v_proj" in k for k in state_dict)
5185
            has_bias = any(".lora_B.bias" in k for k in state_dict)
5186
5187
5188
5189
5190
5191

            if is_i2v_lora:
                return state_dict

            for i in range(num_blocks):
                for o, c in zip(["k_img", "v_img"], ["add_k_proj", "add_v_proj"]):
5192
5193
5194
5195
5196
5197
5198
                    # These keys should exist if the block `i` was part of the T2V LoRA.
                    ref_key_lora_A = f"transformer.blocks.{i}.attn2.to_k.lora_A.weight"
                    ref_key_lora_B = f"transformer.blocks.{i}.attn2.to_k.lora_B.weight"

                    if ref_key_lora_A not in state_dict or ref_key_lora_B not in state_dict:
                        continue

5199
                    state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_A.weight"] = torch.zeros_like(
5200
                        state_dict[f"transformer.blocks.{i}.attn2.to_k.lora_A.weight"], device=target_device
5201
5202
                    )
                    state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_B.weight"] = torch.zeros_like(
5203
                        state_dict[f"transformer.blocks.{i}.attn2.to_k.lora_B.weight"], device=target_device
5204
5205
                    )

5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
                    # If the original LoRA had biases (indicated by has_bias)
                    # AND the specific reference bias key exists for this block.

                    ref_key_lora_B_bias = f"transformer.blocks.{i}.attn2.to_k.lora_B.bias"
                    if has_bias and ref_key_lora_B_bias in state_dict:
                        ref_lora_B_bias_tensor = state_dict[ref_key_lora_B_bias]
                        state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_B.bias"] = torch.zeros_like(
                            ref_lora_B_bias_tensor,
                            device=target_device,
                        )

5217
5218
        return state_dict

Aryan's avatar
Aryan committed
5219
    def load_lora_weights(
5220
5221
5222
5223
5224
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
Aryan's avatar
Aryan committed
5225
5226
5227
5228
5229
5230
5231
5232
5233
5234
5235
5236
5237
5238
5239
5240
5241
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
5242
5243
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
Aryan's avatar
Aryan committed
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
5261
5262
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
5263
5264
5265
5266
5267
        # convert T2V LoRA to I2V LoRA (when loaded to Wan I2V) by adding zeros for the additional (missing) _img layers
        state_dict = self._maybe_expand_t2v_lora_for_i2v(
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            state_dict=state_dict,
        )
Aryan's avatar
Aryan committed
5268
5269
5270
5271
5272
5273
5274
5275
        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
5276
            metadata=metadata,
Aryan's avatar
Aryan committed
5277
5278
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
5279
            hotswap=hotswap,
Aryan's avatar
Aryan committed
5280
5281
5282
5283
5284
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->WanTransformer3DModel
    def load_lora_into_transformer(
5285
5286
5287
5288
5289
5290
5291
5292
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
Aryan's avatar
Aryan committed
5293
5294
5295
5296
5297
5298
5299
5300
5301
5302
5303
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`WanTransformer3DModel`):
                The Transformer model to load the LoRA layers into.
Aryan's avatar
Aryan committed
5304
5305
5306
5307
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
5308
5309
5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
5385
5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
5398
5399
5400
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
5426
5427
5428
5429
5430
5431
5432
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443
5444
5445
5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
5477
5478
5479
5480
5481
5482
5483
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
5546
5547
5548
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
5591
5592
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
5627
5628
5629
5630
5631
5632
5633
5634
5635
5636
5637
5638
5639
5640
5641
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
5682
5683
5684
5685
5686
5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
            metadata=metadata,
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
            hotswap=hotswap,
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
        transformer_lora_adapter_metadata: Optional[dict] = None,
    ):
        r"""
        Save the LoRA parameters corresponding to the transformer.

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
        """
        state_dict = {}
        lora_adapter_metadata = {}

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
            lora_adapter_metadata=lora_adapter_metadata,
        )

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
    def fuse_lora(
        self,
        components: List[str] = ["transformer"],
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
        )

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
        super().unfuse_lora(components=components, **kwargs)


class SkyReelsV2LoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`SkyReelsV2Transformer3DModel`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    # Copied from diffusers.loaders.lora_pipeline.WanLoraLoaderMixin.lora_state_dict
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}

        state_dict, metadata = _fetch_state_dict(
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )
        if any(k.startswith("diffusion_model.") for k in state_dict):
            state_dict = _convert_non_diffusers_wan_lora_to_diffusers(state_dict)
        elif any(k.startswith("lora_unet_") for k in state_dict):
            state_dict = _convert_musubi_wan_lora_to_diffusers(state_dict)

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.WanLoraLoaderMixin._maybe_expand_t2v_lora_for_i2v
    def _maybe_expand_t2v_lora_for_i2v(
        cls,
        transformer: torch.nn.Module,
        state_dict,
    ):
        if transformer.config.image_dim is None:
            return state_dict

        target_device = transformer.device

        if any(k.startswith("transformer.blocks.") for k in state_dict):
            num_blocks = len({k.split("blocks.")[1].split(".")[0] for k in state_dict if "blocks." in k})
            is_i2v_lora = any("add_k_proj" in k for k in state_dict) and any("add_v_proj" in k for k in state_dict)
            has_bias = any(".lora_B.bias" in k for k in state_dict)

            if is_i2v_lora:
                return state_dict

            for i in range(num_blocks):
                for o, c in zip(["k_img", "v_img"], ["add_k_proj", "add_v_proj"]):
                    # These keys should exist if the block `i` was part of the T2V LoRA.
                    ref_key_lora_A = f"transformer.blocks.{i}.attn2.to_k.lora_A.weight"
                    ref_key_lora_B = f"transformer.blocks.{i}.attn2.to_k.lora_B.weight"

                    if ref_key_lora_A not in state_dict or ref_key_lora_B not in state_dict:
                        continue

                    state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_A.weight"] = torch.zeros_like(
                        state_dict[f"transformer.blocks.{i}.attn2.to_k.lora_A.weight"], device=target_device
                    )
                    state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_B.weight"] = torch.zeros_like(
                        state_dict[f"transformer.blocks.{i}.attn2.to_k.lora_B.weight"], device=target_device
                    )

                    # If the original LoRA had biases (indicated by has_bias)
                    # AND the specific reference bias key exists for this block.

                    ref_key_lora_B_bias = f"transformer.blocks.{i}.attn2.to_k.lora_B.bias"
                    if has_bias and ref_key_lora_B_bias in state_dict:
                        ref_lora_B_bias_tensor = state_dict[ref_key_lora_B_bias]
                        state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_B.bias"] = torch.zeros_like(
                            ref_lora_B_bias_tensor,
                            device=target_device,
                        )

        return state_dict

    # Copied from diffusers.loaders.lora_pipeline.WanLoraLoaderMixin.load_lora_weights
    def load_lora_weights(
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
        # convert T2V LoRA to I2V LoRA (when loaded to Wan I2V) by adding zeros for the additional (missing) _img layers
        state_dict = self._maybe_expand_t2v_lora_for_i2v(
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            state_dict=state_dict,
        )
        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
            metadata=metadata,
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
            hotswap=hotswap,
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->SkyReelsV2Transformer3DModel
    def load_lora_into_transformer(
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`SkyReelsV2Transformer3DModel`):
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
Aryan's avatar
Aryan committed
5706
5707
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
5708
5709
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
5710
5711
5712
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
Aryan's avatar
Aryan committed
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
5725
            metadata=metadata,
Aryan's avatar
Aryan committed
5726
5727
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
5728
            hotswap=hotswap,
Aryan's avatar
Aryan committed
5729
5730
5731
5732
5733
5734
5735
5736
5737
5738
5739
5740
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
5741
        transformer_lora_adapter_metadata: Optional[dict] = None,
Aryan's avatar
Aryan committed
5742
5743
    ):
        r"""
5744
        Save the LoRA parameters corresponding to the transformer.
Aryan's avatar
Aryan committed
5745
5746
5747
5748
5749
5750
5751
5752
5753
5754
5755
5756
5757
5758
5759
5760

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
5761
5762
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
Aryan's avatar
Aryan committed
5763
5764
        """
        state_dict = {}
5765
        lora_adapter_metadata = {}
Aryan's avatar
Aryan committed
5766
5767
5768
5769

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

5770
5771
5772
5773
5774
5775
        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )
Aryan's avatar
Aryan committed
5776
5777
5778
5779
5780
5781
5782
5783
5784

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
5785
            lora_adapter_metadata=lora_adapter_metadata,
Aryan's avatar
Aryan committed
5786
5787
5788
5789
5790
5791
5792
5793
5794
5795
5796
5797
5798
5799
5800
5801
5802
5803
5804
5805
5806
5807
5808
5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
5820
5821
5822
5823
5824
5825
5826
5827
5828
        )

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
    def fuse_lora(
        self,
        components: List[str] = ["transformer"],
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
5829
5830
5831
5832
5833
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
Aryan's avatar
Aryan committed
5834
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
        )

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
5852
        super().unfuse_lora(components=components, **kwargs)
Aryan's avatar
Aryan committed
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
5872
5873
5874
5875
5876
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913


class CogView4LoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`WanTransformer3DModel`]. Specific to [`CogView4Pipeline`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.lora_state_dict
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
5914
5915
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
Aryan's avatar
Aryan committed
5916
5917
5918
5919
5920
5921
5922
5923
5924
5925
5926
5927
5928

        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
5929
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
Aryan's avatar
Aryan committed
5930
5931
5932
5933
5934
5935

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

5936
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
Aryan's avatar
Aryan committed
5937

5938
        state_dict, metadata = _fetch_state_dict(
Aryan's avatar
Aryan committed
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
5956
5957
5958
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

5959
5960
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
Aryan's avatar
Aryan committed
5961
5962
5963

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
    def load_lora_weights(
5964
5965
5966
5967
5968
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
Aryan's avatar
Aryan committed
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
5983
5984
5985
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
5986
5987
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
Aryan's avatar
Aryan committed
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
6005
6006
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
Aryan's avatar
Aryan committed
6007
6008
6009
6010
6011
6012
6013
6014
6015

        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
6016
            metadata=metadata,
Aryan's avatar
Aryan committed
6017
6018
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
6019
            hotswap=hotswap,
Aryan's avatar
Aryan committed
6020
6021
6022
6023
6024
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->CogView4Transformer2DModel
    def load_lora_into_transformer(
6025
6026
6027
6028
6029
6030
6031
6032
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
Aryan's avatar
Aryan committed
6033
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`CogView4Transformer2DModel`):
                The Transformer model to load the LoRA layers into.
Aryan's avatar
Aryan committed
6044
6045
6046
6047
6048
6049
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
6050
6051
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
6052
6053
6054
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
Aryan's avatar
Aryan committed
6055
6056
6057
6058
6059
6060
6061
6062
6063
6064
6065
6066
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
6067
            metadata=metadata,
Aryan's avatar
Aryan committed
6068
6069
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
6070
            hotswap=hotswap,
Aryan's avatar
Aryan committed
6071
6072
6073
6074
6075
6076
6077
6078
6079
6080
6081
6082
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
6083
        transformer_lora_adapter_metadata: Optional[dict] = None,
Aryan's avatar
Aryan committed
6084
6085
    ):
        r"""
6086
        Save the LoRA parameters corresponding to the transformer.
Aryan's avatar
Aryan committed
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
6103
6104
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
Aryan's avatar
Aryan committed
6105
6106
        """
        state_dict = {}
6107
        lora_adapter_metadata = {}
Aryan's avatar
Aryan committed
6108
6109
6110
6111

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

6112
6113
6114
6115
6116
6117
        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )
Aryan's avatar
Aryan committed
6118
6119
6120
6121
6122
6123
6124
6125
6126

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
6127
            lora_adapter_metadata=lora_adapter_metadata,
Aryan's avatar
Aryan committed
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
6156
6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
        )

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
    def fuse_lora(
        self,
        components: List[str] = ["transformer"],
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
6171
6172
6173
6174
6175
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
Aryan's avatar
Aryan committed
6176
6177
6178
6179
6180
6181
6182
6183
6184
6185
6186
6187
6188
6189
6190
6191
6192
6193
        )

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
6194
        super().unfuse_lora(components=components, **kwargs)
Aryan's avatar
Aryan committed
6195
6196


6197
6198
6199
6200
6201
6202
6203
6204
6205
6206
6207
6208
6209
6210
6211
6212
6213
6214
6215
6216
6217
6218
6219
6220
6221
6222
6223
6224
6225
6226
6227
6228
6229
6230
6231
6232
6233
6234
6235
6236
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
class HiDreamImageLoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`HiDreamImageTransformer2DModel`]. Specific to [`HiDreamImagePipeline`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
6255
6256
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
6267
6268
        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
6269
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
6270
6271
6272
6273
6274
6275

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

6276
        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
6277

6278
        state_dict, metadata = _fetch_state_dict(
6279
6280
6281
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
6292
6293
6294
6295
6296
6297
6298
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

6299
6300
6301
6302
        is_non_diffusers_format = any("diffusion_model" in k for k in state_dict)
        if is_non_diffusers_format:
            state_dict = _convert_non_diffusers_hidream_lora_to_diffusers(state_dict)

6303
6304
        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out
6305
6306
6307
6308
6309
6310
6311
6312
6313
6314
6315
6316
6317
6318
6319
6320
6321
6322
6323
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
6339
6340
6341
6342
6343
6344
6345
6346
6347
6348

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
    def load_lora_weights(
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
6349
6350
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
6351
6352
6353
6354
6355
6356
6357
6358
6359

        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
6360
            metadata=metadata,
6361
6362
6363
6364
6365
6366
6367
6368
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
            hotswap=hotswap,
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->HiDreamImageTransformer2DModel
    def load_lora_into_transformer(
6369
6370
6371
6372
6373
6374
6375
6376
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
6377
6378
6379
6380
6381
6382
6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`HiDreamImageTransformer2DModel`):
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
6396
6397
6398
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
6399
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
6411
            metadata=metadata,
6412
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
6423
6424
6425
6426
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
            hotswap=hotswap,
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
6427
        transformer_lora_adapter_metadata: Optional[dict] = None,
6428
6429
    ):
        r"""
6430
        Save the LoRA parameters corresponding to the transformer.
6431
6432
6433
6434
6435
6436
6437
6438
6439
6440
6441
6442
6443
6444
6445
6446

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
6447
6448
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
6449
6450
        """
        state_dict = {}
6451
        lora_adapter_metadata = {}
6452
6453
6454
6455

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

6456
6457
6458
6459
6460
6461
        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )
6462
6463
6464
6465
6466
6467
6468
6469
6470

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
6471
            lora_adapter_metadata=lora_adapter_metadata,
6472
6473
6474
6475
6476
6477
6478
6479
6480
6481
6482
6483
6484
6485
6486
6487
6488
6489
6490
6491
6492
6493
6494
6495
6496
6497
6498
6499
6500
6501
6502
6503
6504
6505
6506
6507
6508
6509
6510
6511
6512
6513
6514
6515
6516
6517
6518
6519
6520
6521
6522
6523
6524
6525
6526
6527
6528
6529
6530
6531
6532
6533
6534
6535
6536
6537
6538
6539
6540
        )

    # Copied from diffusers.loaders.lora_pipeline.SanaLoraLoaderMixin.fuse_lora
    def fuse_lora(
        self,
        components: List[str] = ["transformer"],
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
        )

    # Copied from diffusers.loaders.lora_pipeline.SanaLoraLoaderMixin.unfuse_lora
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
        super().unfuse_lora(components=components, **kwargs)


6541
6542
6543
6544
6545
6546
6547
6548
6549
6550
6551
6552
6553
6554
6555
6556
6557
6558
6559
6560
6561
6562
6563
6564
6565
6566
6567
6568
6569
6570
6571
6572
6573
6574
6575
6576
6577
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
6594
6595
6596
6597
6598
6599
6600
6601
6602
6603
6604
6605
6606
6607
6608
6609
6610
6611
6612
6613
6614
6615
6616
6617
6618
6619
6620
6621
6622
6623
6624
6625
6626
6627
6628
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
6639
6640
6641
6642
6643
6644
6645
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657
6658
6659
6660
6661
6662
6663
6664
6665
6666
6667
6668
6669
6670
6671
6672
6673
6674
6675
6676
6677
6678
6679
6680
6681
6682
6683
6684
6685
6686
6687
6688
6689
6690
6691
6692
6693
6694
6695
6696
6697
6698
6699
6700
6701
6702
6703
6704
6705
6706
6707
6708
6709
6710
6711
6712
6713
6714
6715
6716
6717
6718
6719
6720
6721
6722
6723
6724
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
6742
6743
6744
6745
6746
6747
6748
6749
6750
6751
6752
6753
6754
6755
6756
6757
6758
6759
6760
6761
6762
6763
6764
6765
6766
6767
6768
6769
6770
6771
6772
6773
6774
6775
6776
6777
6778
6779
6780
6781
6782
6783
6784
6785
6786
6787
6788
6789
6790
6791
6792
6793
6794
6795
6796
6797
6798
6799
6800
6801
6802
6803
6804
6805
6806
6807
6808
6809
6810
6811
6812
6813
6814
6815
6816
6817
6818
6819
6820
6821
6822
6823
6824
6825
6826
6827
6828
6829
6830
6831
6832
6833
6834
6835
6836
6837
6838
6839
6840
6841
6842
6843
6844
6845
6846
6847
6848
6849
6850
6851
6852
6853
6854
6855
6856
6857
6858
6859
6860
6861
6862
6863
6864
6865
6866
6867
6868
6869
6870
6871
6872
6873
6874
6875
6876
6877
6878
6879
6880
6881
6882
class QwenImageLoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`QwenImageTransformer2DModel`]. Specific to [`QwenImagePipeline`].
    """

    _lora_loadable_modules = ["transformer"]
    transformer_name = TRANSFORMER_NAME

    @classmethod
    @validate_hf_hub_args
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.lora_state_dict
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        **kwargs,
    ):
        r"""
        Return state dict for lora weights and the network alphas.

        <Tip warning={true}>

        We support loading A1111 formatted LoRA checkpoints in a limited capacity.

        This function is experimental and might change in the future.

        </Tip>

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                Can be either:

                    - 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`].
                    - A [torch state
                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).

            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
            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*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            local_files_only (`bool`, *optional*, defaults to `False`):
                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.
            token (`str` or *bool*, *optional*):
                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.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
            return_lora_metadata (`bool`, *optional*, defaults to False):
                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.

        """
        # Load the main state dict first which has the LoRA layers for either of
        # transformer and text encoder or both.
        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)
        weight_name = kwargs.pop("weight_name", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
        return_lora_metadata = kwargs.pop("return_lora_metadata", False)

        allow_pickle = False
        if use_safetensors is None:
            use_safetensors = True
            allow_pickle = True

        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}

        state_dict, metadata = _fetch_state_dict(
            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
            weight_name=weight_name,
            use_safetensors=use_safetensors,
            local_files_only=local_files_only,
            cache_dir=cache_dir,
            force_download=force_download,
            proxies=proxies,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            allow_pickle=allow_pickle,
        )

        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
        if is_dora_scale_present:
            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
            logger.warning(warn_msg)
            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

        out = (state_dict, metadata) if return_lora_metadata else state_dict
        return out

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
    def load_lora_weights(
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
        adapter_name: Optional[str] = None,
        hotswap: bool = False,
        **kwargs,
    ):
        """
        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
        dict is loaded into `self.transformer`.

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        if not USE_PEFT_BACKEND:
            raise ValueError("PEFT backend is required for this method.")

        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # if a dict is passed, copy it instead of modifying it inplace
        if isinstance(pretrained_model_name_or_path_or_dict, dict):
            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()

        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
        kwargs["return_lora_metadata"] = True
        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)

        is_correct_format = all("lora" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_transformer(
            state_dict,
            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
            adapter_name=adapter_name,
            metadata=metadata,
            _pipeline=self,
            low_cpu_mem_usage=low_cpu_mem_usage,
            hotswap=hotswap,
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->QwenImageTransformer2DModel
    def load_lora_into_transformer(
        cls,
        state_dict,
        transformer,
        adapter_name=None,
        _pipeline=None,
        low_cpu_mem_usage=False,
        hotswap: bool = False,
        metadata=None,
    ):
        """
        This will load the LoRA layers specified in `state_dict` into `transformer`.

        Parameters:
            state_dict (`dict`):
                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                encoder lora layers.
            transformer (`QwenImageTransformer2DModel`):
                The Transformer model to load the LoRA layers into.
            adapter_name (`str`, *optional*):
                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
                `default_{i}` where i is the total number of adapters being loaded.
            low_cpu_mem_usage (`bool`, *optional*):
                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
                weights.
            hotswap (`bool`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
            metadata (`dict`):
                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
                from the state dict.
        """
        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
            raise ValueError(
                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
            )

        # Load the layers corresponding to transformer.
        logger.info(f"Loading {cls.transformer_name}.")
        transformer.load_lora_adapter(
            state_dict,
            network_alphas=None,
            adapter_name=adapter_name,
            metadata=metadata,
            _pipeline=_pipeline,
            low_cpu_mem_usage=low_cpu_mem_usage,
            hotswap=hotswap,
        )

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
        transformer_lora_adapter_metadata: Optional[dict] = None,
    ):
        r"""
        Save the LoRA parameters corresponding to the transformer.

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save LoRA parameters to. Will be created if it doesn't exist.
            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            is_main_process (`bool`, *optional*, defaults to `True`):
                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.
            save_function (`Callable`):
                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
                `DIFFUSERS_SAVE_MODE`.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
            transformer_lora_adapter_metadata:
                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
        """
        state_dict = {}
        lora_adapter_metadata = {}

        if not transformer_lora_layers:
            raise ValueError("You must pass `transformer_lora_layers`.")

        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))

        if transformer_lora_adapter_metadata is not None:
            lora_adapter_metadata.update(
                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
            )

        # Save the model
        cls.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
            lora_adapter_metadata=lora_adapter_metadata,
        )

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
    def fuse_lora(
        self,
        components: List[str] = ["transformer"],
        lora_scale: float = 1.0,
        safe_fusing: bool = False,
        adapter_names: Optional[List[str]] = None,
        **kwargs,
    ):
        r"""
        Fuses the LoRA parameters into the original parameters of the corresponding blocks.

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
            lora_scale (`float`, defaults to 1.0):
                Controls how much to influence the outputs with the LoRA parameters.
            safe_fusing (`bool`, defaults to `False`):
                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
            adapter_names (`List[str]`, *optional*):
                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

        Example:

        ```py
        from diffusers import DiffusionPipeline
        import torch

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
        ).to("cuda")
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        super().fuse_lora(
            components=components,
            lora_scale=lora_scale,
            safe_fusing=safe_fusing,
            adapter_names=adapter_names,
            **kwargs,
        )

    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
        r"""
        Reverses the effect of
        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).

        <Tip warning={true}>

        This is an experimental API.

        </Tip>

        Args:
            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
        """
        super().unfuse_lora(components=components, **kwargs)


6883
6884
6885
6886
6887
class LoraLoaderMixin(StableDiffusionLoraLoaderMixin):
    def __init__(self, *args, **kwargs):
        deprecation_message = "LoraLoaderMixin is deprecated and this will be removed in a future version. Please use `StableDiffusionLoraLoaderMixin`, instead."
        deprecate("LoraLoaderMixin", "1.0.0", deprecation_message)
        super().__init__(*args, **kwargs)