Skip to content

LoRA

LoRA is a fast and lightweight training method that inserts and trains a significantly smaller number of parameters instead of all the model parameters. This produces a smaller file (~100 MBs) and makes it easier to quickly train a model to learn a new concept. LoRA weights are typically loaded into the denoiser, text encoder or both. The denoiser usually corresponds to a UNet (UNet2DConditionModel, for example) or a Transformer (SD3Transformer2DModel, for example). There are several classes for loading LoRA weights:

  • StableDiffusionLoraLoaderMixin provides functions for loading and unloading, fusing and unfusing, enabling and disabling, and more functions for managing LoRA weights. This class can be used with any model.
  • StableDiffusionXLLoraLoaderMixin is a Stable Diffusion (SDXL) version of the StableDiffusionLoraLoaderMixin class for loading and saving LoRA weights. It can only be used with the SDXL model.
  • SD3LoraLoaderMixin provides similar functions for Stable Diffusion 3
  • LoraBaseMixin provides a base class with several utility methods to fuse, unfuse, unload, LoRAs and more.

Tip

To learn more about how to load LoRA weights, see the LoRA loading guide.

mindone.diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin

Bases: LoraBaseMixin

Load LoRA layers into Stable Diffusion [UNet2DConditionModel] and CLIPTextModel.

Source code in mindone/diffusers/loaders/lora_pipeline.py
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
class StableDiffusionLoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into Stable Diffusion [`UNet2DConditionModel`] and
    [`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel).
    """

    _lora_loadable_modules = ["unet", "text_encoder"]
    unet_name = UNET_NAME
    text_encoder_name = TEXT_ENCODER_NAME

    def load_lora_weights(
        self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.Tensor]], adapter_name=None, **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`.

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

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

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

        Parameters:
            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
            kwargs (`dict`, *optional*):
                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.
        """
        # 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.
        state_dict, network_alphas = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)

        is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
        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,
            _pipeline=self,
        )
        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,
        )

    @classmethod
    @validate_hf_hub_args
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.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 MindSpore 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.
            weight_name (`str`, *optional*, defaults to None):
                Name of the serialized state dict file.
        """
        # 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)
        unet_config = kwargs.pop("unet_config", None)
        use_safetensors = kwargs.pop("use_safetensors", None)

        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 = cls._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,
        )

        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)

        return state_dict, network_alphas

    @classmethod
    def load_lora_into_unet(cls, state_dict, network_alphas, unet, adapter_name=None, _pipeline=None):
        """
        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.
        """
        # 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.
        keys = list(state_dict.keys())
        only_text_encoder = all(key.startswith(cls.text_encoder_name) for key in keys)
        if not only_text_encoder:
            # Load the layers corresponding to UNet.
            logger.info(f"Loading {cls.unet_name}.")
            unet.load_attn_procs(
                state_dict, network_alphas=network_alphas, adapter_name=adapter_name, _pipeline=_pipeline
            )

    @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,
    ):
        """
        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]`):
                See `LoRALinearLayer` for more details.
            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.
        """
        from mindone.diffusers._peft import LoraConfig

        # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
        # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
        # their prefixes.
        keys = list(state_dict.keys())
        prefix = cls.text_encoder_name if prefix is None else prefix

        # Safe prefix to check with.
        if any(cls.text_encoder_name in key for key in keys):
            # Load the layers corresponding to text encoder and make necessary adjustments.
            text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
            text_encoder_lora_state_dict = {
                k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
            }

            if len(text_encoder_lora_state_dict) > 0:
                logger.info(f"Loading {prefix}.")
                rank = {}
                text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)

                # convert state dict
                text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)

                for name, _ in text_encoder_attn_modules(text_encoder):
                    for module in ("out_proj", "q_proj", "k_proj", "v_proj"):
                        rank_key = f"{name}.{module}.lora_B.weight"
                        if rank_key not in text_encoder_lora_state_dict:
                            continue
                        rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

                for name, _ in text_encoder_mlp_modules(text_encoder):
                    for module in ("fc1", "fc2"):
                        rank_key = f"{name}.{module}.lora_B.weight"
                        if rank_key not in text_encoder_lora_state_dict:
                            continue
                        rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

                if network_alphas is not None:
                    alpha_keys = [
                        k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
                    ]
                    network_alphas = {
                        k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
                    }

                lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
                if "use_dora" in lora_config_kwargs:
                    if lora_config_kwargs["use_dora"]:
                        if is_peft_version("<", "0.9.0"):
                            raise ValueError(
                                "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
                            )
                    else:
                        if is_peft_version("<", "0.9.0"):
                            lora_config_kwargs.pop("use_dora")
                lora_config = LoraConfig(**lora_config_kwargs)

                # adapter_name
                if adapter_name is None:
                    adapter_name = get_adapter_name(text_encoder)

                # inject LoRA layers and load the state dict
                # in transformers we automatically check whether the adapter name is already in use or not
                text_encoder.load_adapter(
                    adapter_name=adapter_name,
                    adapter_state_dict=text_encoder_lora_state_dict,
                    peft_config=lora_config,
                )

                # scale LoRA layers with `lora_scale`
                scale_lora_layers(text_encoder, weight=lora_scale)

                text_encoder.to(dtype=text_encoder.dtype)

    @classmethod
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        unet_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = None,
        text_encoder_lora_layers: Dict[str, nn.Cell] = 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.
            unet_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]`):
                State dict of the LoRA layers corresponding to the `unet`.
            text_encoder_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.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 `mindspore.save_checkpoint` 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 MindSpore way.
        """
        state_dict = {}

        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`.")

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

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

    def fuse_lora(
        self,
        components: List[str] = ["unet", "text_encoder"],
        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 mindone.diffusers import DiffusionPipeline
        import mindspore

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
        )
        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
        )

    def unfuse_lora(self, components: List[str] = ["unet", "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_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.
        """
        super().unfuse_lora(components=components)

mindone.diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.fuse_lora(components=['unet', 'text_encoder'], lora_scale=1.0, safe_fusing=False, adapter_names=None, **kwargs)

Fuses the LoRA parameters into the original parameters of the corresponding blocks.

This is an experimental API.

PARAMETER DESCRIPTION
components

(List[str]): List of LoRA-injectable components to fuse the LoRAs into.

TYPE: List[str] DEFAULT: ['unet', 'text_encoder']

lora_scale

Controls how much to influence the outputs with the LoRA parameters.

TYPE: `float`, defaults to 1.0 DEFAULT: 1.0

safe_fusing

Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.

TYPE: `bool`, defaults to `False` DEFAULT: False

adapter_names

Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

TYPE: `List[str]`, *optional* DEFAULT: None

from mindone.diffusers import DiffusionPipeline
import mindspore

pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
)
pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
pipeline.fuse_lora(lora_scale=0.7)
Source code in mindone/diffusers/loaders/lora_pipeline.py
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
def fuse_lora(
    self,
    components: List[str] = ["unet", "text_encoder"],
    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 mindone.diffusers import DiffusionPipeline
    import mindspore

    pipeline = DiffusionPipeline.from_pretrained(
        "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
    )
    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
    )

mindone.diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder(state_dict, network_alphas, text_encoder, prefix=None, lora_scale=1.0, adapter_name=None, _pipeline=None) classmethod

This will load the LoRA layers specified in state_dict into text_encoder

PARAMETER DESCRIPTION
state_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.

TYPE: `dict`

network_alphas

See LoRALinearLayer for more details.

TYPE: `Dict[str, float]`

text_encoder

The text encoder model to load the LoRA layers into.

TYPE: `CLIPTextModel`

prefix

Expected prefix of the text_encoder in the state_dict.

TYPE: `str` DEFAULT: None

lora_scale

How much to scale the output of the lora linear layer before it is added with the output of the regular lora layer.

TYPE: `float` DEFAULT: 1.0

adapter_name

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.

TYPE: `str`, *optional* DEFAULT: None

Source code in mindone/diffusers/loaders/lora_pipeline.py
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
@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,
):
    """
    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]`):
            See `LoRALinearLayer` for more details.
        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.
    """
    from mindone.diffusers._peft import LoraConfig

    # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
    # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
    # their prefixes.
    keys = list(state_dict.keys())
    prefix = cls.text_encoder_name if prefix is None else prefix

    # Safe prefix to check with.
    if any(cls.text_encoder_name in key for key in keys):
        # Load the layers corresponding to text encoder and make necessary adjustments.
        text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
        text_encoder_lora_state_dict = {
            k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
        }

        if len(text_encoder_lora_state_dict) > 0:
            logger.info(f"Loading {prefix}.")
            rank = {}
            text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)

            # convert state dict
            text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)

            for name, _ in text_encoder_attn_modules(text_encoder):
                for module in ("out_proj", "q_proj", "k_proj", "v_proj"):
                    rank_key = f"{name}.{module}.lora_B.weight"
                    if rank_key not in text_encoder_lora_state_dict:
                        continue
                    rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

            for name, _ in text_encoder_mlp_modules(text_encoder):
                for module in ("fc1", "fc2"):
                    rank_key = f"{name}.{module}.lora_B.weight"
                    if rank_key not in text_encoder_lora_state_dict:
                        continue
                    rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

            if network_alphas is not None:
                alpha_keys = [
                    k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
                ]
                network_alphas = {
                    k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
                }

            lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
            if "use_dora" in lora_config_kwargs:
                if lora_config_kwargs["use_dora"]:
                    if is_peft_version("<", "0.9.0"):
                        raise ValueError(
                            "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
                        )
                else:
                    if is_peft_version("<", "0.9.0"):
                        lora_config_kwargs.pop("use_dora")
            lora_config = LoraConfig(**lora_config_kwargs)

            # adapter_name
            if adapter_name is None:
                adapter_name = get_adapter_name(text_encoder)

            # inject LoRA layers and load the state dict
            # in transformers we automatically check whether the adapter name is already in use or not
            text_encoder.load_adapter(
                adapter_name=adapter_name,
                adapter_state_dict=text_encoder_lora_state_dict,
                peft_config=lora_config,
            )

            # scale LoRA layers with `lora_scale`
            scale_lora_layers(text_encoder, weight=lora_scale)

            text_encoder.to(dtype=text_encoder.dtype)

mindone.diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_unet(state_dict, network_alphas, unet, adapter_name=None, _pipeline=None) classmethod

This will load the LoRA layers specified in state_dict into unet.

PARAMETER DESCRIPTION
state_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.

TYPE: `dict`

network_alphas

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.

TYPE: `Dict[str, float]`

unet

The UNet model to load the LoRA layers into.

TYPE: `UNet2DConditionModel`

adapter_name

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.

TYPE: `str`, *optional* DEFAULT: None

Source code in mindone/diffusers/loaders/lora_pipeline.py
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
@classmethod
def load_lora_into_unet(cls, state_dict, network_alphas, unet, adapter_name=None, _pipeline=None):
    """
    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.
    """
    # 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.
    keys = list(state_dict.keys())
    only_text_encoder = all(key.startswith(cls.text_encoder_name) for key in keys)
    if not only_text_encoder:
        # Load the layers corresponding to UNet.
        logger.info(f"Loading {cls.unet_name}.")
        unet.load_attn_procs(
            state_dict, network_alphas=network_alphas, adapter_name=adapter_name, _pipeline=_pipeline
        )

mindone.diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_weights(pretrained_model_name_or_path_or_dict, adapter_name=None, **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.

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

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

See [~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder] for more details on how the state dict is loaded into self.text_encoder.

PARAMETER DESCRIPTION
pretrained_model_name_or_path_or_dict

See [~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict].

TYPE: `str` or `os.PathLike` or `dict`

kwargs

See [~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict].

TYPE: `dict`, *optional* DEFAULT: {}

adapter_name

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.

TYPE: `str`, *optional* DEFAULT: None

Source code in mindone/diffusers/loaders/lora_pipeline.py
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
def load_lora_weights(
    self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.Tensor]], adapter_name=None, **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`.

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

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

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

    Parameters:
        pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
            See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        kwargs (`dict`, *optional*):
            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.
    """
    # 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.
    state_dict, network_alphas = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)

    is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
    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,
        _pipeline=self,
    )
    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,
    )

mindone.diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs) classmethod

Return state dict for lora weights and the network alphas.

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

This function is experimental and might change in the future.

PARAMETER DESCRIPTION
pretrained_model_name_or_path_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 MindSpore state dict

TYPE: `str` or `os.PathLike` or `dict`

cache_dir

Path to a directory where a downloaded pretrained model configuration is cached if the standard cache is not used.

TYPE: `Union[str, os.PathLike]`, *optional*

force_download

Whether or not to force the (re-)download of the model weights and configuration files, overriding the cached versions if they exist.

TYPE: `bool`, *optional*, defaults to `False`

proxies

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.

TYPE: `Dict[str, str]`, *optional*

local_files_only

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.

TYPE: `bool`, *optional*, defaults to `False`

token

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.

TYPE: `str` or *bool*, *optional*

revision

The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier allowed by Git.

TYPE: `str`, *optional*, defaults to `"main"`

subfolder

The subfolder location of a model file within a larger model repository on the Hub or locally.

TYPE: `str`, *optional*, defaults to `""`

weight_name

Name of the serialized state dict file.

TYPE: `str`, *optional*, defaults to None

Source code in mindone/diffusers/loaders/lora_pipeline.py
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
@classmethod
@validate_hf_hub_args
def lora_state_dict(
    cls,
    pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.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 MindSpore 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.
        weight_name (`str`, *optional*, defaults to None):
            Name of the serialized state dict file.
    """
    # 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)
    unet_config = kwargs.pop("unet_config", None)
    use_safetensors = kwargs.pop("use_safetensors", None)

    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 = cls._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,
    )

    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)

    return state_dict, network_alphas

mindone.diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.save_lora_weights(save_directory, unet_lora_layers=None, text_encoder_lora_layers=None, is_main_process=True, weight_name=None, save_function=None, safe_serialization=True) classmethod

Save the LoRA parameters corresponding to the UNet and text encoder.

PARAMETER DESCRIPTION
save_directory

Directory to save LoRA parameters to. Will be created if it doesn't exist.

TYPE: `str` or `os.PathLike`

unet_lora_layers

State dict of the LoRA layers corresponding to the unet.

TYPE: `Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]` DEFAULT: None

text_encoder_lora_layers

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.

TYPE: `Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]` DEFAULT: None

is_main_process

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.

TYPE: `bool`, *optional*, defaults to `True` DEFAULT: True

save_function

The function to use to save the state dictionary. Useful during distributed training when you need to replace mindspore.save_checkpoint with another method. Can be configured with the environment variable DIFFUSERS_SAVE_MODE.

TYPE: `Callable` DEFAULT: None

safe_serialization

Whether to save the model using safetensors or the traditional MindSpore way.

TYPE: `bool`, *optional*, defaults to `True` DEFAULT: True

Source code in mindone/diffusers/loaders/lora_pipeline.py
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
@classmethod
def save_lora_weights(
    cls,
    save_directory: Union[str, os.PathLike],
    unet_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = None,
    text_encoder_lora_layers: Dict[str, nn.Cell] = 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.
        unet_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]`):
            State dict of the LoRA layers corresponding to the `unet`.
        text_encoder_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.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 `mindspore.save_checkpoint` 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 MindSpore way.
    """
    state_dict = {}

    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`.")

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

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

mindone.diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.unfuse_lora(components=['unet', 'text_encoder'], **kwargs)

Reverses the effect of pipe.fuse_lora().

This is an experimental API.

PARAMETER DESCRIPTION
components

List of LoRA-injectable components to unfuse LoRA from.

TYPE: `List[str]` DEFAULT: ['unet', 'text_encoder']

unfuse_unet

Whether to unfuse the UNet LoRA parameters.

TYPE: `bool`, defaults to `True`

unfuse_text_encoder

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.

TYPE: `bool`, defaults to `True`

Source code in mindone/diffusers/loaders/lora_pipeline.py
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
def unfuse_lora(self, components: List[str] = ["unet", "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_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.
    """
    super().unfuse_lora(components=components)

mindone.diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin

Bases: LoraBaseMixin

Load LoRA layers into Stable Diffusion XL [UNet2DConditionModel], CLIPTextModel, and CLIPTextModelWithProjection.

Source code in mindone/diffusers/loaders/lora_pipeline.py
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
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
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
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).
    """

    _lora_loadable_modules = ["unet", "text_encoder", "text_encoder_2"]
    unet_name = UNET_NAME
    text_encoder_name = TEXT_ENCODER_NAME

    def load_lora_weights(
        self,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.Tensor]],
        adapter_name: Optional[str] = None,
        **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`.

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

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

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

        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.
            kwargs (`dict`, *optional*):
                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
        """
        # 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.

        # 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.
        state_dict, network_alphas = self.lora_state_dict(
            pretrained_model_name_or_path_or_dict,
            unet_config=self.unet.config,
            **kwargs,
        )
        is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
        if not is_correct_format:
            raise ValueError("Invalid LoRA checkpoint.")

        self.load_lora_into_unet(
            state_dict, network_alphas=network_alphas, unet=self.unet, adapter_name=adapter_name, _pipeline=self
        )
        text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
        if len(text_encoder_state_dict) > 0:
            self.load_lora_into_text_encoder(
                text_encoder_state_dict,
                network_alphas=network_alphas,
                text_encoder=self.text_encoder,
                prefix="text_encoder",
                lora_scale=self.lora_scale,
                adapter_name=adapter_name,
                _pipeline=self,
            )

        text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
        if len(text_encoder_2_state_dict) > 0:
            self.load_lora_into_text_encoder(
                text_encoder_2_state_dict,
                network_alphas=network_alphas,
                text_encoder=self.text_encoder_2,
                prefix="text_encoder_2",
                lora_scale=self.lora_scale,
                adapter_name=adapter_name,
                _pipeline=self,
            )

    @classmethod
    @validate_hf_hub_args
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.lora_state_dict
    def lora_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.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 MindSpore 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.
            weight_name (`str`, *optional*, defaults to None):
                Name of the serialized state dict file.
        """
        # 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)
        unet_config = kwargs.pop("unet_config", None)
        use_safetensors = kwargs.pop("use_safetensors", None)

        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 = cls._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,
        )

        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)

        return state_dict, network_alphas

    @classmethod
    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_unet
    def load_lora_into_unet(cls, state_dict, network_alphas, unet, adapter_name=None, _pipeline=None):
        """
        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.
        """
        # 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.
        keys = list(state_dict.keys())
        only_text_encoder = all(key.startswith(cls.text_encoder_name) for key in keys)
        if not only_text_encoder:
            # Load the layers corresponding to UNet.
            logger.info(f"Loading {cls.unet_name}.")
            unet.load_attn_procs(
                state_dict, network_alphas=network_alphas, adapter_name=adapter_name, _pipeline=_pipeline
            )

    @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,
    ):
        """
        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]`):
                See `LoRALinearLayer` for more details.
            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.
        """
        from mindone.diffusers._peft import LoraConfig

        # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
        # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
        # their prefixes.
        keys = list(state_dict.keys())
        prefix = cls.text_encoder_name if prefix is None else prefix

        # Safe prefix to check with.
        if any(cls.text_encoder_name in key for key in keys):
            # Load the layers corresponding to text encoder and make necessary adjustments.
            text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
            text_encoder_lora_state_dict = {
                k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
            }

            if len(text_encoder_lora_state_dict) > 0:
                logger.info(f"Loading {prefix}.")
                rank = {}
                text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)

                # convert state dict
                text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)

                for name, _ in text_encoder_attn_modules(text_encoder):
                    for module in ("out_proj", "q_proj", "k_proj", "v_proj"):
                        rank_key = f"{name}.{module}.lora_B.weight"
                        if rank_key not in text_encoder_lora_state_dict:
                            continue
                        rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

                for name, _ in text_encoder_mlp_modules(text_encoder):
                    for module in ("fc1", "fc2"):
                        rank_key = f"{name}.{module}.lora_B.weight"
                        if rank_key not in text_encoder_lora_state_dict:
                            continue
                        rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

                if network_alphas is not None:
                    alpha_keys = [
                        k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
                    ]
                    network_alphas = {
                        k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
                    }

                lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
                if "use_dora" in lora_config_kwargs:
                    if lora_config_kwargs["use_dora"]:
                        if is_peft_version("<", "0.9.0"):
                            raise ValueError(
                                "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
                            )
                    else:
                        if is_peft_version("<", "0.9.0"):
                            lora_config_kwargs.pop("use_dora")
                lora_config = LoraConfig(**lora_config_kwargs)

                # adapter_name
                if adapter_name is None:
                    adapter_name = get_adapter_name(text_encoder)

                # inject LoRA layers and load the state dict
                # in transformers we automatically check whether the adapter name is already in use or not
                text_encoder.load_adapter(
                    adapter_name=adapter_name,
                    adapter_state_dict=text_encoder_lora_state_dict,
                    peft_config=lora_config,
                )

                # scale LoRA layers with `lora_scale`
                scale_lora_layers(text_encoder, weight=lora_scale)

                text_encoder.to(dtype=text_encoder.dtype)

    @classmethod
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        unet_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = None,
        text_encoder_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = None,
        text_encoder_2_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = 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.
            unet_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]`):
                State dict of the LoRA layers corresponding to the `unet`.
            text_encoder_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.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, nn.Cell]` or `Dict[str, ms.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 `mindspore.save_checkpoint` 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 MindSpore way.
        """
        state_dict = {}

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

        if unet_lora_layers:
            state_dict.update(cls.pack_weights(unet_lora_layers, "unet"))

        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"))

        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,
        )

    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 mindone.diffusers import DiffusionPipeline
        import mindspore

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
        )
        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
        )

    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.
        """
        super().unfuse_lora(components=components)

mindone.diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.fuse_lora(components=['unet', 'text_encoder', 'text_encoder_2'], lora_scale=1.0, safe_fusing=False, adapter_names=None, **kwargs)

Fuses the LoRA parameters into the original parameters of the corresponding blocks.

This is an experimental API.

PARAMETER DESCRIPTION
components

(List[str]): List of LoRA-injectable components to fuse the LoRAs into.

TYPE: List[str] DEFAULT: ['unet', 'text_encoder', 'text_encoder_2']

lora_scale

Controls how much to influence the outputs with the LoRA parameters.

TYPE: `float`, defaults to 1.0 DEFAULT: 1.0

safe_fusing

Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.

TYPE: `bool`, defaults to `False` DEFAULT: False

adapter_names

Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

TYPE: `List[str]`, *optional* DEFAULT: None

from mindone.diffusers import DiffusionPipeline
import mindspore

pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
)
pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
pipeline.fuse_lora(lora_scale=0.7)
Source code in mindone/diffusers/loaders/lora_pipeline.py
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
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 mindone.diffusers import DiffusionPipeline
    import mindspore

    pipeline = DiffusionPipeline.from_pretrained(
        "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
    )
    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
    )

mindone.diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.load_lora_into_text_encoder(state_dict, network_alphas, text_encoder, prefix=None, lora_scale=1.0, adapter_name=None, _pipeline=None) classmethod

This will load the LoRA layers specified in state_dict into text_encoder

PARAMETER DESCRIPTION
state_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.

TYPE: `dict`

network_alphas

See LoRALinearLayer for more details.

TYPE: `Dict[str, float]`

text_encoder

The text encoder model to load the LoRA layers into.

TYPE: `CLIPTextModel`

prefix

Expected prefix of the text_encoder in the state_dict.

TYPE: `str` DEFAULT: None

lora_scale

How much to scale the output of the lora linear layer before it is added with the output of the regular lora layer.

TYPE: `float` DEFAULT: 1.0

adapter_name

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.

TYPE: `str`, *optional* DEFAULT: None

Source code in mindone/diffusers/loaders/lora_pipeline.py
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
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
@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,
):
    """
    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]`):
            See `LoRALinearLayer` for more details.
        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.
    """
    from mindone.diffusers._peft import LoraConfig

    # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
    # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
    # their prefixes.
    keys = list(state_dict.keys())
    prefix = cls.text_encoder_name if prefix is None else prefix

    # Safe prefix to check with.
    if any(cls.text_encoder_name in key for key in keys):
        # Load the layers corresponding to text encoder and make necessary adjustments.
        text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
        text_encoder_lora_state_dict = {
            k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
        }

        if len(text_encoder_lora_state_dict) > 0:
            logger.info(f"Loading {prefix}.")
            rank = {}
            text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)

            # convert state dict
            text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)

            for name, _ in text_encoder_attn_modules(text_encoder):
                for module in ("out_proj", "q_proj", "k_proj", "v_proj"):
                    rank_key = f"{name}.{module}.lora_B.weight"
                    if rank_key not in text_encoder_lora_state_dict:
                        continue
                    rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

            for name, _ in text_encoder_mlp_modules(text_encoder):
                for module in ("fc1", "fc2"):
                    rank_key = f"{name}.{module}.lora_B.weight"
                    if rank_key not in text_encoder_lora_state_dict:
                        continue
                    rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

            if network_alphas is not None:
                alpha_keys = [
                    k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
                ]
                network_alphas = {
                    k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
                }

            lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
            if "use_dora" in lora_config_kwargs:
                if lora_config_kwargs["use_dora"]:
                    if is_peft_version("<", "0.9.0"):
                        raise ValueError(
                            "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
                        )
                else:
                    if is_peft_version("<", "0.9.0"):
                        lora_config_kwargs.pop("use_dora")
            lora_config = LoraConfig(**lora_config_kwargs)

            # adapter_name
            if adapter_name is None:
                adapter_name = get_adapter_name(text_encoder)

            # inject LoRA layers and load the state dict
            # in transformers we automatically check whether the adapter name is already in use or not
            text_encoder.load_adapter(
                adapter_name=adapter_name,
                adapter_state_dict=text_encoder_lora_state_dict,
                peft_config=lora_config,
            )

            # scale LoRA layers with `lora_scale`
            scale_lora_layers(text_encoder, weight=lora_scale)

            text_encoder.to(dtype=text_encoder.dtype)

mindone.diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.load_lora_into_unet(state_dict, network_alphas, unet, adapter_name=None, _pipeline=None) classmethod

This will load the LoRA layers specified in state_dict into unet.

PARAMETER DESCRIPTION
state_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.

TYPE: `dict`

network_alphas

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.

TYPE: `Dict[str, float]`

unet

The UNet model to load the LoRA layers into.

TYPE: `UNet2DConditionModel`

adapter_name

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.

TYPE: `str`, *optional* DEFAULT: None

Source code in mindone/diffusers/loaders/lora_pipeline.py
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
@classmethod
# Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_unet
def load_lora_into_unet(cls, state_dict, network_alphas, unet, adapter_name=None, _pipeline=None):
    """
    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.
    """
    # 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.
    keys = list(state_dict.keys())
    only_text_encoder = all(key.startswith(cls.text_encoder_name) for key in keys)
    if not only_text_encoder:
        # Load the layers corresponding to UNet.
        logger.info(f"Loading {cls.unet_name}.")
        unet.load_attn_procs(
            state_dict, network_alphas=network_alphas, adapter_name=adapter_name, _pipeline=_pipeline
        )

mindone.diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.load_lora_weights(pretrained_model_name_or_path_or_dict, adapter_name=None, **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.

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

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

See [~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder] for more details on how the state dict is loaded into self.text_encoder.

PARAMETER DESCRIPTION
pretrained_model_name_or_path_or_dict

See [~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict].

TYPE: `str` or `os.PathLike` or `dict`

adapter_name

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.

TYPE: `str`, *optional* DEFAULT: None

kwargs

See [~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict].

TYPE: `dict`, *optional* DEFAULT: {}

Source code in mindone/diffusers/loaders/lora_pipeline.py
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
def load_lora_weights(
    self,
    pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.Tensor]],
    adapter_name: Optional[str] = None,
    **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`.

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

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

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

    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.
        kwargs (`dict`, *optional*):
            See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
    """
    # 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.

    # 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.
    state_dict, network_alphas = self.lora_state_dict(
        pretrained_model_name_or_path_or_dict,
        unet_config=self.unet.config,
        **kwargs,
    )
    is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
    if not is_correct_format:
        raise ValueError("Invalid LoRA checkpoint.")

    self.load_lora_into_unet(
        state_dict, network_alphas=network_alphas, unet=self.unet, adapter_name=adapter_name, _pipeline=self
    )
    text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
    if len(text_encoder_state_dict) > 0:
        self.load_lora_into_text_encoder(
            text_encoder_state_dict,
            network_alphas=network_alphas,
            text_encoder=self.text_encoder,
            prefix="text_encoder",
            lora_scale=self.lora_scale,
            adapter_name=adapter_name,
            _pipeline=self,
        )

    text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
    if len(text_encoder_2_state_dict) > 0:
        self.load_lora_into_text_encoder(
            text_encoder_2_state_dict,
            network_alphas=network_alphas,
            text_encoder=self.text_encoder_2,
            prefix="text_encoder_2",
            lora_scale=self.lora_scale,
            adapter_name=adapter_name,
            _pipeline=self,
        )

mindone.diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs) classmethod

Return state dict for lora weights and the network alphas.

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

This function is experimental and might change in the future.

PARAMETER DESCRIPTION
pretrained_model_name_or_path_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 MindSpore state dict.

TYPE: `str` or `os.PathLike` or `dict`

cache_dir

Path to a directory where a downloaded pretrained model configuration is cached if the standard cache is not used.

TYPE: `Union[str, os.PathLike]`, *optional*

force_download

Whether or not to force the (re-)download of the model weights and configuration files, overriding the cached versions if they exist.

TYPE: `bool`, *optional*, defaults to `False`

proxies

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.

TYPE: `Dict[str, str]`, *optional*

local_files_only

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.

TYPE: `bool`, *optional*, defaults to `False`

token

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.

TYPE: `str` or *bool*, *optional*

revision

The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier allowed by Git.

TYPE: `str`, *optional*, defaults to `"main"`

subfolder

The subfolder location of a model file within a larger model repository on the Hub or locally.

TYPE: `str`, *optional*, defaults to `""`

weight_name

Name of the serialized state dict file.

TYPE: `str`, *optional*, defaults to None

Source code in mindone/diffusers/loaders/lora_pipeline.py
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
@classmethod
@validate_hf_hub_args
# Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.lora_state_dict
def lora_state_dict(
    cls,
    pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.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 MindSpore 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.
        weight_name (`str`, *optional*, defaults to None):
            Name of the serialized state dict file.
    """
    # 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)
    unet_config = kwargs.pop("unet_config", None)
    use_safetensors = kwargs.pop("use_safetensors", None)

    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 = cls._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,
    )

    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)

    return state_dict, network_alphas

mindone.diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.save_lora_weights(save_directory, unet_lora_layers=None, text_encoder_lora_layers=None, text_encoder_2_lora_layers=None, is_main_process=True, weight_name=None, save_function=None, safe_serialization=True) classmethod

Save the LoRA parameters corresponding to the UNet and text encoder.

PARAMETER DESCRIPTION
save_directory

Directory to save LoRA parameters to. Will be created if it doesn't exist.

TYPE: `str` or `os.PathLike`

unet_lora_layers

State dict of the LoRA layers corresponding to the unet.

TYPE: `Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]` DEFAULT: None

text_encoder_lora_layers

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.

TYPE: `Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]` DEFAULT: None

text_encoder_2_lora_layers

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.

TYPE: `Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]` DEFAULT: None

is_main_process

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.

TYPE: `bool`, *optional*, defaults to `True` DEFAULT: True

save_function

The function to use to save the state dictionary. Useful during distributed training when you need to replace mindspore.save_checkpoint with another method. Can be configured with the environment variable DIFFUSERS_SAVE_MODE.

TYPE: `Callable` DEFAULT: None

safe_serialization

Whether to save the model using safetensors or the traditional MindSpore way.

TYPE: `bool`, *optional*, defaults to `True` DEFAULT: True

Source code in mindone/diffusers/loaders/lora_pipeline.py
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
@classmethod
def save_lora_weights(
    cls,
    save_directory: Union[str, os.PathLike],
    unet_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = None,
    text_encoder_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = None,
    text_encoder_2_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = 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.
        unet_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]`):
            State dict of the LoRA layers corresponding to the `unet`.
        text_encoder_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.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, nn.Cell]` or `Dict[str, ms.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 `mindspore.save_checkpoint` 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 MindSpore way.
    """
    state_dict = {}

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

    if unet_lora_layers:
        state_dict.update(cls.pack_weights(unet_lora_layers, "unet"))

    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"))

    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,
    )

mindone.diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.unfuse_lora(components=['unet', 'text_encoder', 'text_encoder_2'], **kwargs)

Reverses the effect of pipe.fuse_lora().

This is an experimental API.

PARAMETER DESCRIPTION
components

List of LoRA-injectable components to unfuse LoRA from.

TYPE: `List[str]` DEFAULT: ['unet', 'text_encoder', 'text_encoder_2']

unfuse_unet

Whether to unfuse the UNet LoRA parameters.

TYPE: `bool`, defaults to `True`

unfuse_text_encoder

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.

TYPE: `bool`, defaults to `True`

Source code in mindone/diffusers/loaders/lora_pipeline.py
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
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.
    """
    super().unfuse_lora(components=components)

mindone.diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin

Bases: LoraBaseMixin

Load LoRA layers into [SD3Transformer2DModel], CLIPTextModel, and CLIPTextModelWithProjection.

Specific to [StableDiffusion3Pipeline].

Source code in mindone/diffusers/loaders/lora_pipeline.py
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
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
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
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, ms.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 MindSpore 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.

        """
        # 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)

        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 = cls._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,
        )

        return state_dict

    def load_lora_weights(
        self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.Tensor]], adapter_name=None, **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`.

        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`].
            kwargs (`dict`, *optional*):
                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.
        """
        # 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.
        state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)

        is_correct_format = all("lora" in key or "dora_scale" 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,
            _pipeline=self,
        )

        text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
        if len(text_encoder_state_dict) > 0:
            self.load_lora_into_text_encoder(
                text_encoder_state_dict,
                network_alphas=None,
                text_encoder=self.text_encoder,
                prefix="text_encoder",
                lora_scale=self.lora_scale,
                adapter_name=adapter_name,
                _pipeline=self,
            )

        text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
        if len(text_encoder_2_state_dict) > 0:
            self.load_lora_into_text_encoder(
                text_encoder_2_state_dict,
                network_alphas=None,
                text_encoder=self.text_encoder_2,
                prefix="text_encoder_2",
                lora_scale=self.lora_scale,
                adapter_name=adapter_name,
                _pipeline=self,
            )

    @classmethod
    def load_lora_into_transformer(cls, state_dict, transformer, adapter_name=None, _pipeline=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 (`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.
        """
        from mindone.diffusers._peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict

        keys = list(state_dict.keys())

        transformer_keys = [k for k in keys if k.startswith(cls.transformer_name)]
        state_dict = {
            k.replace(f"{cls.transformer_name}.", ""): v for k, v in state_dict.items() if k in transformer_keys
        }

        if len(state_dict.keys()) > 0:
            # check with first key if is not in peft format
            first_key = next(iter(state_dict.keys()))
            if "lora_A" not in first_key:
                state_dict = convert_unet_state_dict_to_peft(state_dict)

            if adapter_name in getattr(transformer, "peft_config", {}):
                raise ValueError(
                    f"Adapter name {adapter_name} already in use in the transformer - please select a new adapter name."
                )

            rank = {}
            for key, val in state_dict.items():
                if "lora_B" in key:
                    rank[key] = val.shape[1]

            lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=None, peft_state_dict=state_dict)
            if "use_dora" in lora_config_kwargs:
                if lora_config_kwargs["use_dora"] and is_peft_version("<", "0.9.0"):
                    raise ValueError(
                        "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
                    )
                else:
                    lora_config_kwargs.pop("use_dora")
            lora_config = LoraConfig(**lora_config_kwargs)

            # adapter_name
            if adapter_name is None:
                adapter_name = get_adapter_name(transformer)

            inject_adapter_in_model(lora_config, transformer, adapter_name=adapter_name)
            incompatible_keys = set_peft_model_state_dict(transformer, state_dict, adapter_name)

            if incompatible_keys is not None:
                # check only for unexpected keys
                unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
                if unexpected_keys:
                    logger.warning(
                        f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
                        f" {unexpected_keys}. "
                    )

    @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,
    ):
        """
        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]`):
                See `LoRALinearLayer` for more details.
            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.
        """
        from mindone.diffusers._peft import LoraConfig

        # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
        # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
        # their prefixes.
        keys = list(state_dict.keys())
        prefix = cls.text_encoder_name if prefix is None else prefix

        # Safe prefix to check with.
        if any(cls.text_encoder_name in key for key in keys):
            # Load the layers corresponding to text encoder and make necessary adjustments.
            text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
            text_encoder_lora_state_dict = {
                k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
            }

            if len(text_encoder_lora_state_dict) > 0:
                logger.info(f"Loading {prefix}.")
                rank = {}
                text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)

                # convert state dict
                text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)

                for name, _ in text_encoder_attn_modules(text_encoder):
                    for module in ("out_proj", "q_proj", "k_proj", "v_proj"):
                        rank_key = f"{name}.{module}.lora_B.weight"
                        if rank_key not in text_encoder_lora_state_dict:
                            continue
                        rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

                for name, _ in text_encoder_mlp_modules(text_encoder):
                    for module in ("fc1", "fc2"):
                        rank_key = f"{name}.{module}.lora_B.weight"
                        if rank_key not in text_encoder_lora_state_dict:
                            continue
                        rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

                if network_alphas is not None:
                    alpha_keys = [
                        k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
                    ]
                    network_alphas = {
                        k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
                    }

                lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
                if "use_dora" in lora_config_kwargs:
                    if lora_config_kwargs["use_dora"]:
                        if is_peft_version("<", "0.9.0"):
                            raise ValueError(
                                "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
                            )
                    else:
                        if is_peft_version("<", "0.9.0"):
                            lora_config_kwargs.pop("use_dora")
                lora_config = LoraConfig(**lora_config_kwargs)

                # adapter_name
                if adapter_name is None:
                    adapter_name = get_adapter_name(text_encoder)

                # inject LoRA layers and load the state dict
                # in transformers we automatically check whether the adapter name is already in use or not
                text_encoder.load_adapter(
                    adapter_name=adapter_name,
                    adapter_state_dict=text_encoder_lora_state_dict,
                    peft_config=lora_config,
                )

                # scale LoRA layers with `lora_scale`
                scale_lora_layers(text_encoder, weight=lora_scale)

                text_encoder.to(dtype=text_encoder.dtype)

    @classmethod
    def save_lora_weights(
        cls,
        save_directory: Union[str, os.PathLike],
        transformer_lora_layers: Dict[str, nn.Cell] = None,
        text_encoder_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = None,
        text_encoder_2_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = 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.
            transformer_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]`):
                State dict of the LoRA layers corresponding to the `transformer`.
            text_encoder_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.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, nn.Cell]` or `Dict[str, ms.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 `mindspore.save_checkpoint` 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 MindSpore way.
        """
        state_dict = {}

        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"))

        # 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,
        )

    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 mindone.diffusers import DiffusionPipeline
        import mindspore

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
        )
        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
        )

    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}>

        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.
        """
        super().unfuse_lora(components=components)

mindone.diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.fuse_lora(components=['transformer', 'text_encoder', 'text_encoder_2'], lora_scale=1.0, safe_fusing=False, adapter_names=None, **kwargs)

Fuses the LoRA parameters into the original parameters of the corresponding blocks.

This is an experimental API.

PARAMETER DESCRIPTION
components

(List[str]): List of LoRA-injectable components to fuse the LoRAs into.

TYPE: List[str] DEFAULT: ['transformer', 'text_encoder', 'text_encoder_2']

lora_scale

Controls how much to influence the outputs with the LoRA parameters.

TYPE: `float`, defaults to 1.0 DEFAULT: 1.0

safe_fusing

Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.

TYPE: `bool`, defaults to `False` DEFAULT: False

adapter_names

Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

TYPE: `List[str]`, *optional* DEFAULT: None

from mindone.diffusers import DiffusionPipeline
import mindspore

pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
)
pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
pipeline.fuse_lora(lora_scale=0.7)
Source code in mindone/diffusers/loaders/lora_pipeline.py
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
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 mindone.diffusers import DiffusionPipeline
    import mindspore

    pipeline = DiffusionPipeline.from_pretrained(
        "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
    )
    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
    )

mindone.diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_text_encoder(state_dict, network_alphas, text_encoder, prefix=None, lora_scale=1.0, adapter_name=None, _pipeline=None) classmethod

This will load the LoRA layers specified in state_dict into text_encoder

PARAMETER DESCRIPTION
state_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.

TYPE: `dict`

network_alphas

See LoRALinearLayer for more details.

TYPE: `Dict[str, float]`

text_encoder

The text encoder model to load the LoRA layers into.

TYPE: `CLIPTextModel`

prefix

Expected prefix of the text_encoder in the state_dict.

TYPE: `str` DEFAULT: None

lora_scale

How much to scale the output of the lora linear layer before it is added with the output of the regular lora layer.

TYPE: `float` DEFAULT: 1.0

adapter_name

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.

TYPE: `str`, *optional* DEFAULT: None

Source code in mindone/diffusers/loaders/lora_pipeline.py
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
@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,
):
    """
    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]`):
            See `LoRALinearLayer` for more details.
        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.
    """
    from mindone.diffusers._peft import LoraConfig

    # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
    # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
    # their prefixes.
    keys = list(state_dict.keys())
    prefix = cls.text_encoder_name if prefix is None else prefix

    # Safe prefix to check with.
    if any(cls.text_encoder_name in key for key in keys):
        # Load the layers corresponding to text encoder and make necessary adjustments.
        text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
        text_encoder_lora_state_dict = {
            k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
        }

        if len(text_encoder_lora_state_dict) > 0:
            logger.info(f"Loading {prefix}.")
            rank = {}
            text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)

            # convert state dict
            text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)

            for name, _ in text_encoder_attn_modules(text_encoder):
                for module in ("out_proj", "q_proj", "k_proj", "v_proj"):
                    rank_key = f"{name}.{module}.lora_B.weight"
                    if rank_key not in text_encoder_lora_state_dict:
                        continue
                    rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

            for name, _ in text_encoder_mlp_modules(text_encoder):
                for module in ("fc1", "fc2"):
                    rank_key = f"{name}.{module}.lora_B.weight"
                    if rank_key not in text_encoder_lora_state_dict:
                        continue
                    rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]

            if network_alphas is not None:
                alpha_keys = [
                    k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
                ]
                network_alphas = {
                    k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
                }

            lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
            if "use_dora" in lora_config_kwargs:
                if lora_config_kwargs["use_dora"]:
                    if is_peft_version("<", "0.9.0"):
                        raise ValueError(
                            "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
                        )
                else:
                    if is_peft_version("<", "0.9.0"):
                        lora_config_kwargs.pop("use_dora")
            lora_config = LoraConfig(**lora_config_kwargs)

            # adapter_name
            if adapter_name is None:
                adapter_name = get_adapter_name(text_encoder)

            # inject LoRA layers and load the state dict
            # in transformers we automatically check whether the adapter name is already in use or not
            text_encoder.load_adapter(
                adapter_name=adapter_name,
                adapter_state_dict=text_encoder_lora_state_dict,
                peft_config=lora_config,
            )

            # scale LoRA layers with `lora_scale`
            scale_lora_layers(text_encoder, weight=lora_scale)

            text_encoder.to(dtype=text_encoder.dtype)

mindone.diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer(state_dict, transformer, adapter_name=None, _pipeline=None) classmethod

This will load the LoRA layers specified in state_dict into transformer.

PARAMETER DESCRIPTION
state_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.

TYPE: `dict`

transformer

The Transformer model to load the LoRA layers into.

TYPE: `SD3Transformer2DModel`

adapter_name

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.

TYPE: `str`, *optional* DEFAULT: None

Source code in mindone/diffusers/loaders/lora_pipeline.py
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
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
@classmethod
def load_lora_into_transformer(cls, state_dict, transformer, adapter_name=None, _pipeline=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 (`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.
    """
    from mindone.diffusers._peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict

    keys = list(state_dict.keys())

    transformer_keys = [k for k in keys if k.startswith(cls.transformer_name)]
    state_dict = {
        k.replace(f"{cls.transformer_name}.", ""): v for k, v in state_dict.items() if k in transformer_keys
    }

    if len(state_dict.keys()) > 0:
        # check with first key if is not in peft format
        first_key = next(iter(state_dict.keys()))
        if "lora_A" not in first_key:
            state_dict = convert_unet_state_dict_to_peft(state_dict)

        if adapter_name in getattr(transformer, "peft_config", {}):
            raise ValueError(
                f"Adapter name {adapter_name} already in use in the transformer - please select a new adapter name."
            )

        rank = {}
        for key, val in state_dict.items():
            if "lora_B" in key:
                rank[key] = val.shape[1]

        lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=None, peft_state_dict=state_dict)
        if "use_dora" in lora_config_kwargs:
            if lora_config_kwargs["use_dora"] and is_peft_version("<", "0.9.0"):
                raise ValueError(
                    "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
                )
            else:
                lora_config_kwargs.pop("use_dora")
        lora_config = LoraConfig(**lora_config_kwargs)

        # adapter_name
        if adapter_name is None:
            adapter_name = get_adapter_name(transformer)

        inject_adapter_in_model(lora_config, transformer, adapter_name=adapter_name)
        incompatible_keys = set_peft_model_state_dict(transformer, state_dict, adapter_name)

        if incompatible_keys is not None:
            # check only for unexpected keys
            unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
            if unexpected_keys:
                logger.warning(
                    f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
                    f" {unexpected_keys}. "
                )

mindone.diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_weights(pretrained_model_name_or_path_or_dict, adapter_name=None, **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.

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.

PARAMETER DESCRIPTION
pretrained_model_name_or_path_or_dict

See [~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict].

TYPE: `str` or `os.PathLike` or `dict`

kwargs

See [~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict].

TYPE: `dict`, *optional* DEFAULT: {}

adapter_name

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.

TYPE: `str`, *optional* DEFAULT: None

Source code in mindone/diffusers/loaders/lora_pipeline.py
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
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
def load_lora_weights(
    self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.Tensor]], adapter_name=None, **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`.

    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`].
        kwargs (`dict`, *optional*):
            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.
    """
    # 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.
    state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)

    is_correct_format = all("lora" in key or "dora_scale" 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,
        _pipeline=self,
    )

    text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
    if len(text_encoder_state_dict) > 0:
        self.load_lora_into_text_encoder(
            text_encoder_state_dict,
            network_alphas=None,
            text_encoder=self.text_encoder,
            prefix="text_encoder",
            lora_scale=self.lora_scale,
            adapter_name=adapter_name,
            _pipeline=self,
        )

    text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
    if len(text_encoder_2_state_dict) > 0:
        self.load_lora_into_text_encoder(
            text_encoder_2_state_dict,
            network_alphas=None,
            text_encoder=self.text_encoder_2,
            prefix="text_encoder_2",
            lora_scale=self.lora_scale,
            adapter_name=adapter_name,
            _pipeline=self,
        )

mindone.diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs) classmethod

Return state dict for lora weights and the network alphas.

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

This function is experimental and might change in the future.

PARAMETER DESCRIPTION
pretrained_model_name_or_path_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 MindSpore state dict.

TYPE: `str` or `os.PathLike` or `dict`

cache_dir

Path to a directory where a downloaded pretrained model configuration is cached if the standard cache is not used.

TYPE: `Union[str, os.PathLike]`, *optional*

force_download

Whether or not to force the (re-)download of the model weights and configuration files, overriding the cached versions if they exist.

TYPE: `bool`, *optional*, defaults to `False`

proxies

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.

TYPE: `Dict[str, str]`, *optional*

local_files_only

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.

TYPE: `bool`, *optional*, defaults to `False`

token

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.

TYPE: `str` or *bool*, *optional*

revision

The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier allowed by Git.

TYPE: `str`, *optional*, defaults to `"main"`

subfolder

The subfolder location of a model file within a larger model repository on the Hub or locally.

TYPE: `str`, *optional*, defaults to `""`

Source code in mindone/diffusers/loaders/lora_pipeline.py
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
@classmethod
@validate_hf_hub_args
def lora_state_dict(
    cls,
    pretrained_model_name_or_path_or_dict: Union[str, Dict[str, ms.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 MindSpore 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.

    """
    # 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)

    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 = cls._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,
    )

    return state_dict

mindone.diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.save_lora_weights(save_directory, transformer_lora_layers=None, text_encoder_lora_layers=None, text_encoder_2_lora_layers=None, is_main_process=True, weight_name=None, save_function=None, safe_serialization=True) classmethod

Save the LoRA parameters corresponding to the UNet and text encoder.

PARAMETER DESCRIPTION
save_directory

Directory to save LoRA parameters to. Will be created if it doesn't exist.

TYPE: `str` or `os.PathLike`

transformer_lora_layers

State dict of the LoRA layers corresponding to the transformer.

TYPE: `Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]` DEFAULT: None

text_encoder_lora_layers

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.

TYPE: `Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]` DEFAULT: None

text_encoder_2_lora_layers

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.

TYPE: `Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]` DEFAULT: None

is_main_process

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.

TYPE: `bool`, *optional*, defaults to `True` DEFAULT: True

save_function

The function to use to save the state dictionary. Useful during distributed training when you need to replace mindspore.save_checkpoint with another method. Can be configured with the environment variable DIFFUSERS_SAVE_MODE.

TYPE: `Callable` DEFAULT: None

safe_serialization

Whether to save the model using safetensors or the traditional MindSpore way.

TYPE: `bool`, *optional*, defaults to `True` DEFAULT: True

Source code in mindone/diffusers/loaders/lora_pipeline.py
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
@classmethod
def save_lora_weights(
    cls,
    save_directory: Union[str, os.PathLike],
    transformer_lora_layers: Dict[str, nn.Cell] = None,
    text_encoder_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = None,
    text_encoder_2_lora_layers: Dict[str, Union[nn.Cell, ms.Tensor]] = 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.
        transformer_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.Tensor]`):
            State dict of the LoRA layers corresponding to the `transformer`.
        text_encoder_lora_layers (`Dict[str, nn.Cell]` or `Dict[str, ms.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, nn.Cell]` or `Dict[str, ms.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 `mindspore.save_checkpoint` 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 MindSpore way.
    """
    state_dict = {}

    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"))

    # 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,
    )

mindone.diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.unfuse_lora(components=['transformer', 'text_encoder', 'text_encoder_2'], **kwargs)

Reverses the effect of pipe.fuse_lora().

This is an experimental API.

PARAMETER DESCRIPTION
components

List of LoRA-injectable components to unfuse LoRA from.

TYPE: `List[str]` DEFAULT: ['transformer', 'text_encoder', 'text_encoder_2']

unfuse_unet

Whether to unfuse the UNet LoRA parameters.

TYPE: `bool`, defaults to `True`

unfuse_text_encoder

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.

TYPE: `bool`, defaults to `True`

Source code in mindone/diffusers/loaders/lora_pipeline.py
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
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}>

    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.
    """
    super().unfuse_lora(components=components)

mindone.diffusers.loaders.lora_base.LoraBaseMixin

Utility class for handling LoRAs.

Source code in mindone/diffusers/loaders/lora_base.py
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
class LoraBaseMixin:
    """Utility class for handling LoRAs."""

    _lora_loadable_modules = []
    num_fused_loras = 0

    def load_lora_weights(self, **kwargs):
        raise NotImplementedError("`load_lora_weights()` is not implemented.")

    @classmethod
    def save_lora_weights(cls, **kwargs):
        raise NotImplementedError("`save_lora_weights()` not implemented.")

    @classmethod
    def lora_state_dict(cls, **kwargs):
        raise NotImplementedError("`lora_state_dict()` is not implemented.")

    @classmethod
    def _fetch_state_dict(
        cls,
        pretrained_model_name_or_path_or_dict,
        weight_name,
        use_safetensors,
        local_files_only,
        cache_dir,
        force_download,
        proxies,
        token,
        revision,
        subfolder,
        user_agent,
        allow_pickle,
    ):
        from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE

        model_file = None
        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
            # Let's first try to load .safetensors weights
            if (use_safetensors and weight_name is None) or (
                weight_name is not None and weight_name.endswith(".safetensors")
            ):
                try:
                    # Here we're relaxing the loading check to enable more Inference API
                    # friendliness where sometimes, it's not at all possible to automatically
                    # determine `weight_name`.
                    if weight_name is None:
                        weight_name = cls._best_guess_weight_name(
                            pretrained_model_name_or_path_or_dict,
                            file_extension=".safetensors",
                            local_files_only=local_files_only,
                        )
                    model_file = _get_model_file(
                        pretrained_model_name_or_path_or_dict,
                        weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
                        cache_dir=cache_dir,
                        force_download=force_download,
                        proxies=proxies,
                        local_files_only=local_files_only,
                        token=token,
                        revision=revision,
                        subfolder=subfolder,
                        user_agent=user_agent,
                    )
                    state_dict = load_file(model_file)
                except (IOError, safetensors.SafetensorError) as e:
                    if not allow_pickle:
                        raise e
                    # try loading non-safetensors weights
                    model_file = None
                    pass

            if model_file is None:
                if weight_name is None:
                    weight_name = cls._best_guess_weight_name(
                        pretrained_model_name_or_path_or_dict, file_extension=".bin", local_files_only=local_files_only
                    )
                model_file = _get_model_file(
                    pretrained_model_name_or_path_or_dict,
                    weights_name=weight_name or LORA_WEIGHT_NAME,
                    cache_dir=cache_dir,
                    force_download=force_download,
                    proxies=proxies,
                    local_files_only=local_files_only,
                    token=token,
                    revision=revision,
                    subfolder=subfolder,
                    user_agent=user_agent,
                )
                raise NotImplementedError(
                    f"Only supports deserialization of weights file in safetensors format, but got {model_file}"
                )
        else:
            state_dict = pretrained_model_name_or_path_or_dict

        return state_dict

    @classmethod
    def _best_guess_weight_name(
        cls, pretrained_model_name_or_path_or_dict, file_extension=".safetensors", local_files_only=False
    ):
        from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE

        if local_files_only or HF_HUB_OFFLINE:
            raise ValueError("When using the offline mode, you must specify a `weight_name`.")

        targeted_files = []

        if os.path.isfile(pretrained_model_name_or_path_or_dict):
            return
        elif os.path.isdir(pretrained_model_name_or_path_or_dict):
            targeted_files = [
                f for f in os.listdir(pretrained_model_name_or_path_or_dict) if f.endswith(file_extension)
            ]
        else:
            files_in_repo = model_info(pretrained_model_name_or_path_or_dict).siblings
            targeted_files = [f.rfilename for f in files_in_repo if f.rfilename.endswith(file_extension)]
        if len(targeted_files) == 0:
            return

        # "scheduler" does not correspond to a LoRA checkpoint.
        # "optimizer" does not correspond to a LoRA checkpoint
        # only top-level checkpoints are considered and not the other ones, hence "checkpoint".
        unallowed_substrings = {"scheduler", "optimizer", "checkpoint"}
        targeted_files = list(
            filter(lambda x: all(substring not in x for substring in unallowed_substrings), targeted_files)
        )

        if any(f.endswith(LORA_WEIGHT_NAME) for f in targeted_files):
            targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME), targeted_files))
        elif any(f.endswith(LORA_WEIGHT_NAME_SAFE) for f in targeted_files):
            targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME_SAFE), targeted_files))

        if len(targeted_files) > 1:
            raise ValueError(
                f"Provided path contains more than one weights file in the {file_extension} format. Either specify `weight_name` in `load_lora_weights` or make sure there's only one  `.safetensors` or `.bin` file in  {pretrained_model_name_or_path_or_dict}."  # noqa: E501
            )
        weight_name = targeted_files[0]
        return weight_name

    def unload_lora_weights(self):
        """
        Unloads the LoRA parameters.

        Examples:

        ```python
        >>> # Assuming `pipeline` is already loaded with the LoRA parameters.
        >>> pipeline.unload_lora_weights()
        >>> ...
        ```
        """
        for component in self._lora_loadable_modules:
            model = getattr(self, component, None)
            if model is not None:
                if issubclass(model.__class__, ModelMixin):
                    model.unload_lora()
                elif issubclass(model.__class__, MSPreTrainedModel):
                    _remove_text_encoder_monkey_patch(model)

    def fuse_lora(
        self,
        components: List[str] = [],
        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 mindone.diffusers import DiffusionPipeline
        import mindspore

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
        )
        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
        pipeline.fuse_lora(lora_scale=0.7)
        ```
        """
        if "fuse_unet" in kwargs:
            depr_message = "Passing `fuse_unet` to `fuse_lora()` is deprecated and will be ignored. Please use the `components` argument and provide a list of the components whose LoRAs are to be fused. `fuse_unet` will be removed in a future version."  # noqa: E501
            deprecate(
                "fuse_unet",
                "1.0.0",
                depr_message,
            )
        if "fuse_transformer" in kwargs:
            depr_message = "Passing `fuse_transformer` to `fuse_lora()` is deprecated and will be ignored. Please use the `components` argument and provide a list of the components whose LoRAs are to be fused. `fuse_transformer` will be removed in a future version."  # noqa: E501
            deprecate(
                "fuse_transformer",
                "1.0.0",
                depr_message,
            )
        if "fuse_text_encoder" in kwargs:
            depr_message = "Passing `fuse_text_encoder` to `fuse_lora()` is deprecated and will be ignored. Please use the `components` argument and provide a list of the components whose LoRAs are to be fused. `fuse_text_encoder` will be removed in a future version."  # noqa: E501
            deprecate(
                "fuse_text_encoder",
                "1.0.0",
                depr_message,
            )

        if len(components) == 0:
            raise ValueError("`components` cannot be an empty list.")

        for fuse_component in components:
            if fuse_component not in self._lora_loadable_modules:
                raise ValueError(f"{fuse_component} is not found in {self._lora_loadable_modules=}.")

            model = getattr(self, fuse_component, None)
            if model is not None:
                # check if diffusers model
                if issubclass(model.__class__, ModelMixin):
                    model.fuse_lora(lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names)
                # handle transformers models.
                if issubclass(model.__class__, MSPreTrainedModel):
                    fuse_text_encoder_lora(
                        model, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
                    )

        self.num_fused_loras += 1

    def unfuse_lora(self, components: List[str] = [], **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.
        """
        if "unfuse_unet" in kwargs:
            depr_message = "Passing `unfuse_unet` to `unfuse_lora()` is deprecated and will be ignored. Please use the `components` argument. `unfuse_unet` will be removed in a future version."  # noqa: E501
            deprecate(
                "unfuse_unet",
                "1.0.0",
                depr_message,
            )
        if "unfuse_transformer" in kwargs:
            depr_message = "Passing `unfuse_transformer` to `unfuse_lora()` is deprecated and will be ignored. Please use the `components` argument. `unfuse_transformer` will be removed in a future version."  # noqa: E501
            deprecate(
                "unfuse_transformer",
                "1.0.0",
                depr_message,
            )
        if "unfuse_text_encoder" in kwargs:
            depr_message = "Passing `unfuse_text_encoder` to `unfuse_lora()` is deprecated and will be ignored. Please use the `components` argument. `unfuse_text_encoder` will be removed in a future version."  # noqa: E501
            deprecate(
                "unfuse_text_encoder",
                "1.0.0",
                depr_message,
            )

        if len(components) == 0:
            raise ValueError("`components` cannot be an empty list.")

        for fuse_component in components:
            if fuse_component not in self._lora_loadable_modules:
                raise ValueError(f"{fuse_component} is not found in {self._lora_loadable_modules=}.")

            model = getattr(self, fuse_component, None)
            if model is not None:
                if issubclass(model.__class__, (ModelMixin, MSPreTrainedModel)):
                    for _, module in model.cells_and_names():
                        if isinstance(module, BaseTunerLayer):
                            module.unmerge()

        self.num_fused_loras -= 1

    def set_adapters(
        self,
        adapter_names: Union[List[str], str],
        adapter_weights: Optional[Union[float, Dict, List[float], List[Dict]]] = None,
    ):
        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names

        adapter_weights = copy.deepcopy(adapter_weights)

        # Expand weights into a list, one entry per adapter
        if not isinstance(adapter_weights, list):
            adapter_weights = [adapter_weights] * len(adapter_names)

        if len(adapter_names) != len(adapter_weights):
            raise ValueError(
                f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(adapter_weights)}"
            )

        list_adapters = self.get_list_adapters()  # eg {"unet": ["adapter1", "adapter2"], "text_encoder": ["adapter2"]}
        all_adapters = {
            adapter for adapters in list_adapters.values() for adapter in adapters
        }  # eg ["adapter1", "adapter2"]
        invert_list_adapters = {
            adapter: [part for part, adapters in list_adapters.items() if adapter in adapters]
            for adapter in all_adapters
        }  # eg {"adapter1": ["unet"], "adapter2": ["unet", "text_encoder"]}

        # Decompose weights into weights for denoiser and text encoders.
        _component_adapter_weights = {}
        for component in self._lora_loadable_modules:
            model = getattr(self, component)

            for adapter_name, weights in zip(adapter_names, adapter_weights):
                if isinstance(weights, dict):
                    component_adapter_weights = weights.pop(component, None)

                    if component_adapter_weights is not None and not hasattr(self, component):
                        logger.warning(
                            f"Lora weight dict contains {component} weights but will be ignored because pipeline does not have {component}."
                        )

                    if component_adapter_weights is not None and component not in invert_list_adapters[adapter_name]:
                        logger.warning(
                            (
                                f"Lora weight dict for adapter '{adapter_name}' contains {component},"
                                f"but this will be ignored because {adapter_name} does not contain weights for {component}."
                                f"Valid parts for {adapter_name} are: {invert_list_adapters[adapter_name]}."
                            )
                        )

                else:
                    component_adapter_weights = weights

                _component_adapter_weights.setdefault(component, [])
                _component_adapter_weights[component].append(component_adapter_weights)

            if issubclass(model.__class__, ModelMixin):
                model.set_adapters(adapter_names, _component_adapter_weights[component])
            elif issubclass(model.__class__, MSPreTrainedModel):
                set_adapters_for_text_encoder(adapter_names, model, _component_adapter_weights[component])

    def disable_lora(self):
        for component in self._lora_loadable_modules:
            model = getattr(self, component, None)
            if model is not None:
                if issubclass(model.__class__, ModelMixin):
                    model.disable_lora()
                elif issubclass(model.__class__, MSPreTrainedModel):
                    disable_lora_for_text_encoder(model)

    def enable_lora(self):
        for component in self._lora_loadable_modules:
            model = getattr(self, component, None)
            if model is not None:
                if issubclass(model.__class__, ModelMixin):
                    model.enable_lora()
                elif issubclass(model.__class__, MSPreTrainedModel):
                    enable_lora_for_text_encoder(model)

    def delete_adapters(self, adapter_names: Union[List[str], str]):
        """
        Args:
        Deletes the LoRA layers of `adapter_name` for the unet and text-encoder(s).
            adapter_names (`Union[List[str], str]`):
                The names of the adapter to delete. Can be a single string or a list of strings
        """
        if isinstance(adapter_names, str):
            adapter_names = [adapter_names]

        for component in self._lora_loadable_modules:
            model = getattr(self, component, None)
            if model is not None:
                if issubclass(model.__class__, ModelMixin):
                    model.delete_adapters(adapter_names)
                elif issubclass(model.__class__, MSPreTrainedModel):
                    for adapter_name in adapter_names:
                        delete_adapter_layers(model, adapter_name)

    def get_active_adapters(self) -> List[str]:
        """
        Gets the list of the current active adapters.

        Example:

        ```python
        from mindone.diffusers import DiffusionPipeline

        pipeline = DiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0",
        )
        pipeline.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
        pipeline.get_active_adapters()
        ```
        """
        active_adapters = []

        for component in self._lora_loadable_modules:
            model = getattr(self, component, None)
            if model is not None and issubclass(model.__class__, ModelMixin):
                for _, module in model.cells_and_names():
                    if isinstance(module, BaseTunerLayer):
                        active_adapters = module.active_adapters
                        break

        return active_adapters

    def get_list_adapters(self) -> Dict[str, List[str]]:
        """
        Gets the current list of all available adapters in the pipeline.
        """
        set_adapters = {}

        for component in self._lora_loadable_modules:
            model = getattr(self, component, None)
            if (
                model is not None
                and issubclass(model.__class__, (ModelMixin, MSPreTrainedModel))
                and hasattr(model, "peft_config")
            ):
                set_adapters[component] = list(model.peft_config.keys())

        return set_adapters

    @staticmethod
    def pack_weights(layers, prefix):
        layers_weights = layers.parameters_dict() if isinstance(layers, nn.Cell) else layers
        layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
        return layers_state_dict

    @staticmethod
    def write_lora_layers(
        state_dict: Dict[str, ms.Tensor],
        save_directory: str,
        is_main_process: bool,
        weight_name: str,
        save_function: Callable,
        safe_serialization: bool,
    ):
        from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE

        if os.path.isfile(save_directory):
            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
            return

        if save_function is None:
            if safe_serialization:

                def save_function(weights, filename):
                    return save_file(weights, filename, metadata={"format": "np"})

            else:
                save_function = ms.save_checkpoint

        os.makedirs(save_directory, exist_ok=True)

        if weight_name is None:
            if safe_serialization:
                weight_name = LORA_WEIGHT_NAME_SAFE
            else:
                weight_name = LORA_WEIGHT_NAME

        save_path = Path(save_directory, weight_name).as_posix()
        save_function(state_dict, save_path)
        logger.info(f"Model weights saved in {save_path}")

    @property
    def lora_scale(self) -> float:
        # property function that returns the lora scale which can be set at run time by the pipeline.
        # if _lora_scale has not been set, return 1
        return self._lora_scale if hasattr(self, "_lora_scale") else 1.0

mindone.diffusers.loaders.lora_base.LoraBaseMixin.delete_adapters(adapter_names)

Deletes the LoRA layers of adapter_name for the unet and text-encoder(s). adapter_names (Union[List[str], str]): The names of the adapter to delete. Can be a single string or a list of strings

Source code in mindone/diffusers/loaders/lora_base.py
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
def delete_adapters(self, adapter_names: Union[List[str], str]):
    """
    Args:
    Deletes the LoRA layers of `adapter_name` for the unet and text-encoder(s).
        adapter_names (`Union[List[str], str]`):
            The names of the adapter to delete. Can be a single string or a list of strings
    """
    if isinstance(adapter_names, str):
        adapter_names = [adapter_names]

    for component in self._lora_loadable_modules:
        model = getattr(self, component, None)
        if model is not None:
            if issubclass(model.__class__, ModelMixin):
                model.delete_adapters(adapter_names)
            elif issubclass(model.__class__, MSPreTrainedModel):
                for adapter_name in adapter_names:
                    delete_adapter_layers(model, adapter_name)

mindone.diffusers.loaders.lora_base.LoraBaseMixin.fuse_lora(components=[], lora_scale=1.0, safe_fusing=False, adapter_names=None, **kwargs)

Fuses the LoRA parameters into the original parameters of the corresponding blocks.

This is an experimental API.

PARAMETER DESCRIPTION
components

(List[str]): List of LoRA-injectable components to fuse the LoRAs into.

TYPE: List[str] DEFAULT: []

lora_scale

Controls how much to influence the outputs with the LoRA parameters.

TYPE: `float`, defaults to 1.0 DEFAULT: 1.0

safe_fusing

Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.

TYPE: `bool`, defaults to `False` DEFAULT: False

adapter_names

Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.

TYPE: `List[str]`, *optional* DEFAULT: None

from mindone.diffusers import DiffusionPipeline
import mindspore

pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
)
pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
pipeline.fuse_lora(lora_scale=0.7)
Source code in mindone/diffusers/loaders/lora_base.py
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
def fuse_lora(
    self,
    components: List[str] = [],
    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 mindone.diffusers import DiffusionPipeline
    import mindspore

    pipeline = DiffusionPipeline.from_pretrained(
        "stabilityai/stable-diffusion-xl-base-1.0", mindspore_dtype=mindspore.float16
    )
    pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
    pipeline.fuse_lora(lora_scale=0.7)
    ```
    """
    if "fuse_unet" in kwargs:
        depr_message = "Passing `fuse_unet` to `fuse_lora()` is deprecated and will be ignored. Please use the `components` argument and provide a list of the components whose LoRAs are to be fused. `fuse_unet` will be removed in a future version."  # noqa: E501
        deprecate(
            "fuse_unet",
            "1.0.0",
            depr_message,
        )
    if "fuse_transformer" in kwargs:
        depr_message = "Passing `fuse_transformer` to `fuse_lora()` is deprecated and will be ignored. Please use the `components` argument and provide a list of the components whose LoRAs are to be fused. `fuse_transformer` will be removed in a future version."  # noqa: E501
        deprecate(
            "fuse_transformer",
            "1.0.0",
            depr_message,
        )
    if "fuse_text_encoder" in kwargs:
        depr_message = "Passing `fuse_text_encoder` to `fuse_lora()` is deprecated and will be ignored. Please use the `components` argument and provide a list of the components whose LoRAs are to be fused. `fuse_text_encoder` will be removed in a future version."  # noqa: E501
        deprecate(
            "fuse_text_encoder",
            "1.0.0",
            depr_message,
        )

    if len(components) == 0:
        raise ValueError("`components` cannot be an empty list.")

    for fuse_component in components:
        if fuse_component not in self._lora_loadable_modules:
            raise ValueError(f"{fuse_component} is not found in {self._lora_loadable_modules=}.")

        model = getattr(self, fuse_component, None)
        if model is not None:
            # check if diffusers model
            if issubclass(model.__class__, ModelMixin):
                model.fuse_lora(lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names)
            # handle transformers models.
            if issubclass(model.__class__, MSPreTrainedModel):
                fuse_text_encoder_lora(
                    model, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
                )

    self.num_fused_loras += 1

mindone.diffusers.loaders.lora_base.LoraBaseMixin.get_active_adapters()

Gets the list of the current active adapters.

Example:

from mindone.diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
)
pipeline.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
pipeline.get_active_adapters()
Source code in mindone/diffusers/loaders/lora_base.py
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
def get_active_adapters(self) -> List[str]:
    """
    Gets the list of the current active adapters.

    Example:

    ```python
    from mindone.diffusers import DiffusionPipeline

    pipeline = DiffusionPipeline.from_pretrained(
        "stabilityai/stable-diffusion-xl-base-1.0",
    )
    pipeline.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
    pipeline.get_active_adapters()
    ```
    """
    active_adapters = []

    for component in self._lora_loadable_modules:
        model = getattr(self, component, None)
        if model is not None and issubclass(model.__class__, ModelMixin):
            for _, module in model.cells_and_names():
                if isinstance(module, BaseTunerLayer):
                    active_adapters = module.active_adapters
                    break

    return active_adapters

mindone.diffusers.loaders.lora_base.LoraBaseMixin.get_list_adapters()

Gets the current list of all available adapters in the pipeline.

Source code in mindone/diffusers/loaders/lora_base.py
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
def get_list_adapters(self) -> Dict[str, List[str]]:
    """
    Gets the current list of all available adapters in the pipeline.
    """
    set_adapters = {}

    for component in self._lora_loadable_modules:
        model = getattr(self, component, None)
        if (
            model is not None
            and issubclass(model.__class__, (ModelMixin, MSPreTrainedModel))
            and hasattr(model, "peft_config")
        ):
            set_adapters[component] = list(model.peft_config.keys())

    return set_adapters

mindone.diffusers.loaders.lora_base.LoraBaseMixin.unfuse_lora(components=[], **kwargs)

Reverses the effect of pipe.fuse_lora().

This is an experimental API.

PARAMETER DESCRIPTION
components

List of LoRA-injectable components to unfuse LoRA from.

TYPE: `List[str]` DEFAULT: []

unfuse_unet

Whether to unfuse the UNet LoRA parameters.

TYPE: `bool`, defaults to `True`

unfuse_text_encoder

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.

TYPE: `bool`, defaults to `True`

Source code in mindone/diffusers/loaders/lora_base.py
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
def unfuse_lora(self, components: List[str] = [], **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.
    """
    if "unfuse_unet" in kwargs:
        depr_message = "Passing `unfuse_unet` to `unfuse_lora()` is deprecated and will be ignored. Please use the `components` argument. `unfuse_unet` will be removed in a future version."  # noqa: E501
        deprecate(
            "unfuse_unet",
            "1.0.0",
            depr_message,
        )
    if "unfuse_transformer" in kwargs:
        depr_message = "Passing `unfuse_transformer` to `unfuse_lora()` is deprecated and will be ignored. Please use the `components` argument. `unfuse_transformer` will be removed in a future version."  # noqa: E501
        deprecate(
            "unfuse_transformer",
            "1.0.0",
            depr_message,
        )
    if "unfuse_text_encoder" in kwargs:
        depr_message = "Passing `unfuse_text_encoder` to `unfuse_lora()` is deprecated and will be ignored. Please use the `components` argument. `unfuse_text_encoder` will be removed in a future version."  # noqa: E501
        deprecate(
            "unfuse_text_encoder",
            "1.0.0",
            depr_message,
        )

    if len(components) == 0:
        raise ValueError("`components` cannot be an empty list.")

    for fuse_component in components:
        if fuse_component not in self._lora_loadable_modules:
            raise ValueError(f"{fuse_component} is not found in {self._lora_loadable_modules=}.")

        model = getattr(self, fuse_component, None)
        if model is not None:
            if issubclass(model.__class__, (ModelMixin, MSPreTrainedModel)):
                for _, module in model.cells_and_names():
                    if isinstance(module, BaseTunerLayer):
                        module.unmerge()

    self.num_fused_loras -= 1

mindone.diffusers.loaders.lora_base.LoraBaseMixin.unload_lora_weights()

Unloads the LoRA parameters.

Examples:

>>> # Assuming `pipeline` is already loaded with the LoRA parameters.
>>> pipeline.unload_lora_weights()
>>> ...
Source code in mindone/diffusers/loaders/lora_base.py
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
def unload_lora_weights(self):
    """
    Unloads the LoRA parameters.

    Examples:

    ```python
    >>> # Assuming `pipeline` is already loaded with the LoRA parameters.
    >>> pipeline.unload_lora_weights()
    >>> ...
    ```
    """
    for component in self._lora_loadable_modules:
        model = getattr(self, component, None)
        if model is not None:
            if issubclass(model.__class__, ModelMixin):
                model.unload_lora()
            elif issubclass(model.__class__, MSPreTrainedModel):
                _remove_text_encoder_monkey_patch(model)