|
51 | 51 |
|
52 | 52 | logger = logging.get_logger(__name__) |
53 | 53 |
|
| 54 | +LORA_WEIGHT_NAME = "pytorch_lora_weights.bin" |
| 55 | +LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors" |
| 56 | + |
54 | 57 |
|
55 | 58 | def fuse_text_encoder_lora(text_encoder, lora_scale=1.0, safe_fusing=False, adapter_names=None): |
56 | 59 | """ |
@@ -181,6 +184,119 @@ def _remove_text_encoder_monkey_patch(text_encoder): |
181 | 184 | text_encoder._hf_peft_config_loaded = None |
182 | 185 |
|
183 | 186 |
|
| 187 | +def _fetch_state_dict( |
| 188 | + pretrained_model_name_or_path_or_dict, |
| 189 | + weight_name, |
| 190 | + use_safetensors, |
| 191 | + local_files_only, |
| 192 | + cache_dir, |
| 193 | + force_download, |
| 194 | + proxies, |
| 195 | + token, |
| 196 | + revision, |
| 197 | + subfolder, |
| 198 | + user_agent, |
| 199 | + allow_pickle, |
| 200 | +): |
| 201 | + model_file = None |
| 202 | + if not isinstance(pretrained_model_name_or_path_or_dict, dict): |
| 203 | + # Let's first try to load .safetensors weights |
| 204 | + if (use_safetensors and weight_name is None) or ( |
| 205 | + weight_name is not None and weight_name.endswith(".safetensors") |
| 206 | + ): |
| 207 | + try: |
| 208 | + # Here we're relaxing the loading check to enable more Inference API |
| 209 | + # friendliness where sometimes, it's not at all possible to automatically |
| 210 | + # determine `weight_name`. |
| 211 | + if weight_name is None: |
| 212 | + weight_name = _best_guess_weight_name( |
| 213 | + pretrained_model_name_or_path_or_dict, |
| 214 | + file_extension=".safetensors", |
| 215 | + local_files_only=local_files_only, |
| 216 | + ) |
| 217 | + model_file = _get_model_file( |
| 218 | + pretrained_model_name_or_path_or_dict, |
| 219 | + weights_name=weight_name or LORA_WEIGHT_NAME_SAFE, |
| 220 | + cache_dir=cache_dir, |
| 221 | + force_download=force_download, |
| 222 | + proxies=proxies, |
| 223 | + local_files_only=local_files_only, |
| 224 | + token=token, |
| 225 | + revision=revision, |
| 226 | + subfolder=subfolder, |
| 227 | + user_agent=user_agent, |
| 228 | + ) |
| 229 | + state_dict = safetensors.torch.load_file(model_file, device="cpu") |
| 230 | + except (IOError, safetensors.SafetensorError) as e: |
| 231 | + if not allow_pickle: |
| 232 | + raise e |
| 233 | + # try loading non-safetensors weights |
| 234 | + model_file = None |
| 235 | + pass |
| 236 | + |
| 237 | + if model_file is None: |
| 238 | + if weight_name is None: |
| 239 | + weight_name = _best_guess_weight_name( |
| 240 | + pretrained_model_name_or_path_or_dict, file_extension=".bin", local_files_only=local_files_only |
| 241 | + ) |
| 242 | + model_file = _get_model_file( |
| 243 | + pretrained_model_name_or_path_or_dict, |
| 244 | + weights_name=weight_name or LORA_WEIGHT_NAME, |
| 245 | + cache_dir=cache_dir, |
| 246 | + force_download=force_download, |
| 247 | + proxies=proxies, |
| 248 | + local_files_only=local_files_only, |
| 249 | + token=token, |
| 250 | + revision=revision, |
| 251 | + subfolder=subfolder, |
| 252 | + user_agent=user_agent, |
| 253 | + ) |
| 254 | + state_dict = load_state_dict(model_file) |
| 255 | + else: |
| 256 | + state_dict = pretrained_model_name_or_path_or_dict |
| 257 | + |
| 258 | + return state_dict |
| 259 | + |
| 260 | + |
| 261 | +def _best_guess_weight_name( |
| 262 | + pretrained_model_name_or_path_or_dict, file_extension=".safetensors", local_files_only=False |
| 263 | +): |
| 264 | + if local_files_only or HF_HUB_OFFLINE: |
| 265 | + raise ValueError("When using the offline mode, you must specify a `weight_name`.") |
| 266 | + |
| 267 | + targeted_files = [] |
| 268 | + |
| 269 | + if os.path.isfile(pretrained_model_name_or_path_or_dict): |
| 270 | + return |
| 271 | + elif os.path.isdir(pretrained_model_name_or_path_or_dict): |
| 272 | + targeted_files = [f for f in os.listdir(pretrained_model_name_or_path_or_dict) if f.endswith(file_extension)] |
| 273 | + else: |
| 274 | + files_in_repo = model_info(pretrained_model_name_or_path_or_dict).siblings |
| 275 | + targeted_files = [f.rfilename for f in files_in_repo if f.rfilename.endswith(file_extension)] |
| 276 | + if len(targeted_files) == 0: |
| 277 | + return |
| 278 | + |
| 279 | + # "scheduler" does not correspond to a LoRA checkpoint. |
| 280 | + # "optimizer" does not correspond to a LoRA checkpoint |
| 281 | + # only top-level checkpoints are considered and not the other ones, hence "checkpoint". |
| 282 | + unallowed_substrings = {"scheduler", "optimizer", "checkpoint"} |
| 283 | + targeted_files = list( |
| 284 | + filter(lambda x: all(substring not in x for substring in unallowed_substrings), targeted_files) |
| 285 | + ) |
| 286 | + |
| 287 | + if any(f.endswith(LORA_WEIGHT_NAME) for f in targeted_files): |
| 288 | + targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME), targeted_files)) |
| 289 | + elif any(f.endswith(LORA_WEIGHT_NAME_SAFE) for f in targeted_files): |
| 290 | + targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME_SAFE), targeted_files)) |
| 291 | + |
| 292 | + if len(targeted_files) > 1: |
| 293 | + raise ValueError( |
| 294 | + 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}." |
| 295 | + ) |
| 296 | + weight_name = targeted_files[0] |
| 297 | + return weight_name |
| 298 | + |
| 299 | + |
184 | 300 | class LoraBaseMixin: |
185 | 301 | """Utility class for handling LoRAs.""" |
186 | 302 |
|
@@ -234,124 +350,16 @@ def _optionally_disable_offloading(cls, _pipeline): |
234 | 350 | return (is_model_cpu_offload, is_sequential_cpu_offload) |
235 | 351 |
|
236 | 352 | @classmethod |
237 | | - def _fetch_state_dict( |
238 | | - cls, |
239 | | - pretrained_model_name_or_path_or_dict, |
240 | | - weight_name, |
241 | | - use_safetensors, |
242 | | - local_files_only, |
243 | | - cache_dir, |
244 | | - force_download, |
245 | | - proxies, |
246 | | - token, |
247 | | - revision, |
248 | | - subfolder, |
249 | | - user_agent, |
250 | | - allow_pickle, |
251 | | - ): |
252 | | - from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE |
253 | | - |
254 | | - model_file = None |
255 | | - if not isinstance(pretrained_model_name_or_path_or_dict, dict): |
256 | | - # Let's first try to load .safetensors weights |
257 | | - if (use_safetensors and weight_name is None) or ( |
258 | | - weight_name is not None and weight_name.endswith(".safetensors") |
259 | | - ): |
260 | | - try: |
261 | | - # Here we're relaxing the loading check to enable more Inference API |
262 | | - # friendliness where sometimes, it's not at all possible to automatically |
263 | | - # determine `weight_name`. |
264 | | - if weight_name is None: |
265 | | - weight_name = cls._best_guess_weight_name( |
266 | | - pretrained_model_name_or_path_or_dict, |
267 | | - file_extension=".safetensors", |
268 | | - local_files_only=local_files_only, |
269 | | - ) |
270 | | - model_file = _get_model_file( |
271 | | - pretrained_model_name_or_path_or_dict, |
272 | | - weights_name=weight_name or LORA_WEIGHT_NAME_SAFE, |
273 | | - cache_dir=cache_dir, |
274 | | - force_download=force_download, |
275 | | - proxies=proxies, |
276 | | - local_files_only=local_files_only, |
277 | | - token=token, |
278 | | - revision=revision, |
279 | | - subfolder=subfolder, |
280 | | - user_agent=user_agent, |
281 | | - ) |
282 | | - state_dict = safetensors.torch.load_file(model_file, device="cpu") |
283 | | - except (IOError, safetensors.SafetensorError) as e: |
284 | | - if not allow_pickle: |
285 | | - raise e |
286 | | - # try loading non-safetensors weights |
287 | | - model_file = None |
288 | | - pass |
289 | | - |
290 | | - if model_file is None: |
291 | | - if weight_name is None: |
292 | | - weight_name = cls._best_guess_weight_name( |
293 | | - pretrained_model_name_or_path_or_dict, file_extension=".bin", local_files_only=local_files_only |
294 | | - ) |
295 | | - model_file = _get_model_file( |
296 | | - pretrained_model_name_or_path_or_dict, |
297 | | - weights_name=weight_name or LORA_WEIGHT_NAME, |
298 | | - cache_dir=cache_dir, |
299 | | - force_download=force_download, |
300 | | - proxies=proxies, |
301 | | - local_files_only=local_files_only, |
302 | | - token=token, |
303 | | - revision=revision, |
304 | | - subfolder=subfolder, |
305 | | - user_agent=user_agent, |
306 | | - ) |
307 | | - state_dict = load_state_dict(model_file) |
308 | | - else: |
309 | | - state_dict = pretrained_model_name_or_path_or_dict |
310 | | - |
311 | | - return state_dict |
| 353 | + def _fetch_state_dict(cls, *args, **kwargs): |
| 354 | + deprecation_message = f"Using the `_fetch_state_dict()` method from {cls} has been deprecated and will be removed in a future version. Please use `from diffusers.loaders.lora_base import _fetch_state_dict`." |
| 355 | + deprecate("_fetch_state_dict", "0.35.0", deprecation_message) |
| 356 | + return _fetch_state_dict(*args, **kwargs) |
312 | 357 |
|
313 | 358 | @classmethod |
314 | | - def _best_guess_weight_name( |
315 | | - cls, pretrained_model_name_or_path_or_dict, file_extension=".safetensors", local_files_only=False |
316 | | - ): |
317 | | - from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE |
318 | | - |
319 | | - if local_files_only or HF_HUB_OFFLINE: |
320 | | - raise ValueError("When using the offline mode, you must specify a `weight_name`.") |
321 | | - |
322 | | - targeted_files = [] |
323 | | - |
324 | | - if os.path.isfile(pretrained_model_name_or_path_or_dict): |
325 | | - return |
326 | | - elif os.path.isdir(pretrained_model_name_or_path_or_dict): |
327 | | - targeted_files = [ |
328 | | - f for f in os.listdir(pretrained_model_name_or_path_or_dict) if f.endswith(file_extension) |
329 | | - ] |
330 | | - else: |
331 | | - files_in_repo = model_info(pretrained_model_name_or_path_or_dict).siblings |
332 | | - targeted_files = [f.rfilename for f in files_in_repo if f.rfilename.endswith(file_extension)] |
333 | | - if len(targeted_files) == 0: |
334 | | - return |
335 | | - |
336 | | - # "scheduler" does not correspond to a LoRA checkpoint. |
337 | | - # "optimizer" does not correspond to a LoRA checkpoint |
338 | | - # only top-level checkpoints are considered and not the other ones, hence "checkpoint". |
339 | | - unallowed_substrings = {"scheduler", "optimizer", "checkpoint"} |
340 | | - targeted_files = list( |
341 | | - filter(lambda x: all(substring not in x for substring in unallowed_substrings), targeted_files) |
342 | | - ) |
343 | | - |
344 | | - if any(f.endswith(LORA_WEIGHT_NAME) for f in targeted_files): |
345 | | - targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME), targeted_files)) |
346 | | - elif any(f.endswith(LORA_WEIGHT_NAME_SAFE) for f in targeted_files): |
347 | | - targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME_SAFE), targeted_files)) |
348 | | - |
349 | | - if len(targeted_files) > 1: |
350 | | - raise ValueError( |
351 | | - 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}." |
352 | | - ) |
353 | | - weight_name = targeted_files[0] |
354 | | - return weight_name |
| 359 | + def _best_guess_weight_name(cls, *args, **kwargs): |
| 360 | + deprecation_message = f"Using the `_best_guess_weight_name()` method from {cls} has been deprecated and will be removed in a future version. Please use `from diffusers.loaders.lora_base import _best_guess_weight_name`." |
| 361 | + deprecate("_best_guess_weight_name", "0.35.0", deprecation_message) |
| 362 | + return _best_guess_weight_name(*args, **kwargs) |
355 | 363 |
|
356 | 364 | def unload_lora_weights(self): |
357 | 365 | """ |
@@ -725,8 +733,6 @@ def write_lora_layers( |
725 | 733 | save_function: Callable, |
726 | 734 | safe_serialization: bool, |
727 | 735 | ): |
728 | | - from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE |
729 | | - |
730 | 736 | if os.path.isfile(save_directory): |
731 | 737 | logger.error(f"Provided path ({save_directory}) should be a directory, not a file") |
732 | 738 | return |
|
0 commit comments