Skip to content

Models

๐Ÿค— Diffusers provides pretrained models for popular algorithms and modules to create custom diffusion systems. The primary function of models is to denoise an input sample as modeled by the distribution p{θ}(x{t-1}|x{t})

All models are built from the base ModelMixin class which is a mindspore.nn.Cell providing basic functionality for saving and loading models, locally and from the Hugging Face Hub.

mindone.diffusers.ModelMixin

Bases: Cell, PushToHubMixin

Base class for all models.

[ModelMixin] takes care of storing the model configuration and provides methods for loading, downloading and saving models.

- **config_name** ([`str`]) -- Filename to save a model to when calling [`~models.ModelMixin.save_pretrained`].
Source code in mindone/diffusers/models/modeling_utils.py
 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
 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
 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
 967
 968
 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
class ModelMixin(nn.Cell, PushToHubMixin):
    r"""
    Base class for all models.

    [`ModelMixin`] takes care of storing the model configuration and provides methods for loading, downloading and
    saving models.

        - **config_name** ([`str`]) -- Filename to save a model to when calling [`~models.ModelMixin.save_pretrained`].
    """

    config_name = CONFIG_NAME
    _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]
    _supports_gradient_checkpointing = False
    _keys_to_ignore_on_load_unexpected = None
    _no_split_modules = None
    _keep_in_fp32_modules = None
    _skip_layerwise_casting_patterns = None
    _supports_group_offloading = True

    def __init__(self):
        super().__init__()

        self._gradient_checkpointing_func = None

    def __getattr__(self, name: str) -> Any:
        """The only reason we overwrite `getattr` here is to gracefully deprecate accessing
        config attributes directly. See https://github.com/huggingface/diffusers/pull/3129 We need to overwrite
        __getattr__ here in addition so that we don't trigger `nn.Cell`'s __getattr__':
        https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
        """

        is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name)
        is_attribute = name in self.__dict__

        if is_in_config and not is_attribute:
            deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'unet.config.{name}'."  # noqa: E501
            deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False, stacklevel=3)
            return self._internal_dict[name]

        # call PyTorch's https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
        return super().__getattr__(name)

    @property
    def is_gradient_checkpointing(self) -> bool:
        """
        Whether gradient checkpointing is activated for this model or not.
        """
        return any(hasattr(m, "gradient_checkpointing") and m.gradient_checkpointing for _, m in self.cells_and_names())

    def enable_gradient_checkpointing(self, gradient_checkpointing_func: Optional[Callable] = None) -> None:
        """
        Activates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
        *checkpoint activations* in other frameworks).

        Args:
            gradient_checkpointing_func (`Callable`, *optional*):
                The function to use for gradient checkpointing. If `None`, the default MindSpore checkpointing function
                is used (`mindspore.nn.Cell.recompute_`).
        """
        if not self._supports_gradient_checkpointing:
            raise ValueError(
                f"{self.__class__.__name__} does not support gradient checkpointing. Please make sure to set the boolean attribute "
                f"`_supports_gradient_checkpointing` to `True` in the class definition."
            )

        if gradient_checkpointing_func is None:

            def _gradient_checkpointing_func(module, *args):
                module.recompute_(mode=True)
                return module

            gradient_checkpointing_func = _gradient_checkpointing_func

        self._set_gradient_checkpointing(enable=True)

    def disable_gradient_checkpointing(self) -> None:
        """
        Deactivates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
        *checkpoint activations* in other frameworks).
        """
        if self._supports_gradient_checkpointing:
            self._set_gradient_checkpointing(enable=False)

    def enable_flash_sdp(self, enabled: bool):
        r"""
        .. warning:: This flag is beta and subject to change.

        Enables or disables flash scaled dot product attention.
        """

        # Recursively walk through all the children.
        # Any children which exposes the enable_flash_sdp method
        # gets the message
        def fn_recursive_set_mem_eff(module: nn.Cell):
            if hasattr(module, "enable_flash_sdp"):
                module.enable_flash_sdp(enabled)

            for child in module.cells():
                fn_recursive_set_mem_eff(child)

        for module in self.cells():
            if isinstance(module, nn.Cell):
                fn_recursive_set_mem_eff(module)

    def set_flash_attention_force_cast_dtype(self, force_cast_dtype: Optional[ms.Type]):
        r"""
        Since the flash-attention operator in MindSpore only supports float16 and bfloat16 data types, we need to manually
        set whether to force data type conversion.

        When the attention interface encounters data of an unsupported data type,
        if `force_cast_dtype` is not None, the function will forcibly convert the data to `force_cast_dtype` for computation
        and then restore it to the original data type afterward. If `force_cast_dtype` is None, it will fall back to the
        original attention calculation using mathematical formulas.

        Parameters:
            force_cast_dtype (Optional): The data type to which the input data should be forcibly converted. If None, no forced
            conversion is performed.
        """

        # Recursively walk through all the children.
        # Any children which exposes the set_flash_attention_force_cast_dtype method
        # gets the message
        def fn_recursive_set_mem_eff(module: nn.Cell):
            if hasattr(module, "set_flash_attention_force_cast_dtype"):
                module.set_flash_attention_force_cast_dtype(force_cast_dtype)

            for child in module.cells():
                fn_recursive_set_mem_eff(child)

        for module in self.cells():
            if isinstance(module, nn.Cell):
                fn_recursive_set_mem_eff(module)

    def set_use_memory_efficient_attention_xformers(self, valid: bool, attention_op: Optional[Callable] = None) -> None:
        # Recursively walk through all the children.
        # Any children which exposes the set_use_memory_efficient_attention_xformers method
        # gets the message
        def fn_recursive_set_mem_eff(module: nn.Cell):
            if hasattr(module, "set_use_memory_efficient_attention_xformers"):
                module.set_use_memory_efficient_attention_xformers(valid, attention_op)

            for child in module.cells():
                fn_recursive_set_mem_eff(child)

        for module in self.cells():
            if isinstance(module, nn.Cell):
                fn_recursive_set_mem_eff(module)

    def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None) -> None:
        r"""
        Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).

        When this option is enabled, you should observe lower GPU memory usage and a potential speed up during
        inference. Speed up during training is not guaranteed.

        <Tip warning={true}>

        โš ๏ธ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes
        precedent.

        </Tip>

        Parameters:
            attention_op (`Callable`, *optional*):
                Not supported for now.

        Examples:

        ```py
        >>> import mindspore as ms
        >>> from mindone.diffusers import UNet2DConditionModel

        >>> model = UNet2DConditionModel.from_pretrained(
        ...     "stabilityai/stable-diffusion-2-1", subfolder="unet", mindspore_dtype=ms.float16
        ... )
        >>> model.enable_xformers_memory_efficient_attention()
        ```
        """
        self.set_use_memory_efficient_attention_xformers(True, attention_op)

    def disable_xformers_memory_efficient_attention(self) -> None:
        r"""
        Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
        """
        self.set_use_memory_efficient_attention_xformers(False)

    def enable_layerwise_casting(
        self,
        storage_dtype: ms.Type,
        compute_dtype: Optional[ms.Type] = None,
        skip_modules_pattern: Optional[Tuple[str, ...]] = None,
        skip_modules_classes: Optional[Tuple[Type[nn.Cell], ...]] = None,
        non_blocking: bool = False,
    ) -> None:
        r"""
        Activates layerwise casting for the current model.

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

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

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

            ```python
            >>> from mindone.diffusers import CogVideoXTransformer3DModel

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

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

        Args:
            storage_dtype (`mindspore.Type`):
                The dtype to which the model should be cast for storage.
            compute_dtype (`mindspore.Type`):
                The dtype to which the model weights should be cast during the forward pass.
            skip_modules_pattern (`Tuple[str, ...]`, *optional*):
                A list of patterns to match the names of the modules to skip during the layerwise casting process. If
                set to `None`, default skip patterns are used to ignore certain internal layers of modules and PEFT
                layers.
            skip_modules_classes (`Tuple[Type[nn.Cell], ...]`, *optional*):
                A list of module classes to skip during the layerwise casting process.
            non_blocking (`bool`, *optional*, defaults to `False`):
                If `True`, the weight casting operations are non-blocking.
        """
        raise NotImplementedError("`enable_layerwise_casting` is not yet supported.")

    def enable_group_offload(
        self,
        onload_device: str = "Ascend",
        offload_device: str = "CPU",
        offload_type: str = "block_level",
        num_blocks_per_group: Optional[int] = None,
        non_blocking: bool = False,
        use_stream: bool = False,
        record_stream: bool = False,
        low_cpu_mem_usage=False,
    ) -> None:
        r"""
        Activates group offloading for the current model.

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

        Example:

            ```python
            >>> from mindone.diffusers import CogVideoXTransformer3DModel

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

            >>> transformer.enable_group_offload(
            ...     onload_device="Ascend",
            ...     offload_device="CPU",
            ...     offload_type="leaf_level",
            ...     use_stream=True,
            ... )
            ```
        """
        raise NotImplementedError("`enable_group_offload` is not yet supported.")

    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
        is_main_process: bool = True,
        save_function: Optional[Callable] = None,
        safe_serialization: bool = True,
        variant: Optional[str] = None,
        max_shard_size: Union[int, str] = "10GB",
        push_to_hub: bool = False,
        **kwargs,
    ):
        """
        Save a model and its configuration file to a directory so that it can be reloaded using the
        [`~models.ModelMixin.from_pretrained`] class method.

        Arguments:
            save_directory (`str` or `os.PathLike`):
                Directory to save a model and its configuration file to. Will be created if it doesn't exist.
            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 PyTorch way with `pickle`.
            variant (`str`, *optional*):
                If specified, weights are saved in the format `pytorch_model.<variant>.bin`.
            max_shard_size (`int` or `str`, defaults to `"10GB"`):
                The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size
                lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5GB"`).
                If expressed as an integer, the unit is bytes. Note that this limit will be decreased after a certain
                period of time (starting from Oct 2024) to allow users to upgrade to the latest version of `diffusers`.
                This is to establish a common default size for this argument across different libraries in the Hugging
                Face ecosystem (`transformers`, and `accelerate`, for example).
            push_to_hub (`bool`, *optional*, defaults to `False`):
                Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the
                repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
                namespace).
            kwargs (`Dict[str, Any]`, *optional*):
                Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
        """
        if os.path.isfile(save_directory):
            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
            return

        weights_name = SAFETENSORS_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
        weights_name = _add_variant(weights_name, variant)
        weights_name_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(
            ".safetensors", "{suffix}.safetensors"
        )

        os.makedirs(save_directory, exist_ok=True)

        if push_to_hub:
            commit_message = kwargs.pop("commit_message", None)
            private = kwargs.pop("private", None)
            create_pr = kwargs.pop("create_pr", False)
            token = kwargs.pop("token", None)
            repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
            repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id

        # Only save the model itself if we are using distributed training
        model_to_save = self

        # Attach architecture to the config
        # Save the config
        if is_main_process:
            model_to_save.save_config(save_directory)

        # Save the model
        state_dict = {k: v for k, v in model_to_save.parameters_and_names()}

        # Save the model
        state_dict_split = split_torch_state_dict_into_shards(
            state_dict, max_shard_size=max_shard_size, filename_pattern=weights_name_pattern
        )

        # Clean the folder from a previous save
        if is_main_process:
            for filename in os.listdir(save_directory):
                if filename in state_dict_split.filename_to_tensors.keys():
                    continue
                full_filename = os.path.join(save_directory, filename)
                if not os.path.isfile(full_filename):
                    continue
                weights_without_ext = weights_name_pattern.replace(".bin", "").replace(".safetensors", "")
                weights_without_ext = weights_without_ext.replace("{suffix}", "")
                filename_without_ext = filename.replace(".bin", "").replace(".safetensors", "")
                # make sure that file to be deleted matches format of sharded file, e.g. pytorch_model-00001-of-00005
                if (
                    filename.startswith(weights_without_ext)
                    and _REGEX_SHARD.fullmatch(filename_without_ext) is not None
                ):
                    os.remove(full_filename)

        for filename, tensors in state_dict_split.filename_to_tensors.items():
            shard = {tensor: state_dict[tensor].contiguous() for tensor in tensors}
            filepath = os.path.join(save_directory, filename)
            if safe_serialization:
                # At some point we will need to deal better with save_function (used for TPU and other distributed
                # joyfulness), but for now this enough.
                safe_save_file(shard, filepath, metadata={"format": "np"})
            else:
                ms.save_checkpoint(shard, filepath)

        if state_dict_split.is_sharded:
            index = {
                "metadata": state_dict_split.metadata,
                "weight_map": state_dict_split.tensor_to_filename,
            }
            save_index_file = SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
            save_index_file = os.path.join(save_directory, _add_variant(save_index_file, variant))
            # Save the index as well
            with open(save_index_file, "w", encoding="utf-8") as f:
                content = json.dumps(index, indent=2, sort_keys=True) + "\n"
                f.write(content)
            logger.info(
                f"The model is bigger than the maximum size per checkpoint ({max_shard_size}) and is going to be "
                f"split in {len(state_dict_split.filename_to_tensors)} checkpoint shards. You can find where each parameters has been saved in the "
                f"index located at {save_index_file}."
            )
        else:
            path_to_weights = os.path.join(save_directory, weights_name)
            logger.info(f"Model weights saved in {path_to_weights}")

        if push_to_hub:
            # Create a new empty model card and eventually tag it
            model_card = load_or_create_model_card(repo_id, token=token)
            model_card = populate_model_card(model_card)
            model_card.save(Path(save_directory, "README.md").as_posix())

            self._upload_folder(
                save_directory,
                repo_id,
                token=token,
                commit_message=commit_message,
                create_pr=create_pr,
            )

    @classmethod
    @validate_hf_hub_args
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs) -> Self:
        r"""
        Instantiate a pretrained PyTorch model from a pretrained model configuration.

        The model is set in evaluation mode - `model.eval()` - by default, and dropout modules are deactivated. To
        train the model, set it back in training mode with `model.train()`.

        Parameters:
            pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
                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`].

            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.
            mindspore_dtype (`str` or `mindspore.Type`, *optional*):
                Override the default `mindspore.Type` and load the model with another dtype. If `"auto"` is passed, the
                dtype is automatically derived from the model's weights.
            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.
            output_loading_info (`bool`, *optional*, defaults to `False`):
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
            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.
            from_flax (`bool`, *optional*, defaults to `False`):
                Load the model weights from a Flax checkpoint save file.
            subfolder (`str`, *optional*, defaults to `""`):
                The subfolder location of a model file within a larger model repository on the Hub or locally.
            mirror (`str`, *optional*):
                Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                information.
            variant (`str`, *optional*):
                Load weights from a specified `variant` filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.
            use_safetensors (`bool`, *optional*, defaults to `None`):
                If set to `None`, the `safetensors` weights are downloaded if they're available **and** if the
                `safetensors` library is installed. If set to `True`, the model is forcibly loaded from `safetensors`
                weights. If set to `False`, `safetensors` weights are not loaded.
            disable_mmap ('bool', *optional*, defaults to 'False'):
                Whether to disable mmap when loading a Safetensors model. This option can perform better when the model
                is on a network mount or hard drive, which may not handle the seeky-ness of mmap very well.

        <Tip>

        To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with
        `huggingface-cli login`. You can also activate the special
        ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use this method in a
        firewalled environment.

        </Tip>

        Example:

        ```py
        from mindone.diffusers import UNet2DConditionModel

        unet = UNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="unet")
        ```

        If you get the error message below, you need to finetune the weights for your downstream task:

        ```bash
        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at
        runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
        - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
        You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
        ```
        """
        cache_dir = kwargs.pop("cache_dir", None)
        ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
        force_download = kwargs.pop("force_download", False)
        from_flax = kwargs.pop("from_flax", False)
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
        local_files_only = kwargs.pop("local_files_only", None)
        token = kwargs.pop("token", None)
        revision = kwargs.pop("revision", None)
        mindspore_dtype = kwargs.pop("mindspore_dtype", None)
        subfolder = kwargs.pop("subfolder", None)
        variant = kwargs.pop("variant", None)
        use_safetensors = kwargs.pop("use_safetensors", None)
        dduf_entries: Optional[Dict[str, DDUFEntry]] = kwargs.pop("dduf_entries", None)
        disable_mmap = kwargs.pop("disable_mmap", False)

        if mindspore_dtype is not None and not isinstance(mindspore_dtype, ms.Type):
            mindspore_dtype = ms.float32
            logger.warning(
                f"Passed `mindspore_dtype` {mindspore_dtype} is not a `ms.Type`. Defaulting to `ms.float32`."
            )

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

        user_agent = {
            "diffusers": __version__,
            "file_type": "model",
            "framework": "pytorch",
        }
        unused_kwargs = {}

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

        # load config
        config, unused_kwargs, commit_hash = cls.load_config(
            config_path,
            cache_dir=cache_dir,
            return_unused_kwargs=True,
            return_commit_hash=True,
            force_download=force_download,
            proxies=proxies,
            local_files_only=local_files_only,
            token=token,
            revision=revision,
            subfolder=subfolder,
            user_agent=user_agent,
            dduf_entries=dduf_entries,
            **kwargs,
        )
        # no in-place modification of the original config.
        config = copy.deepcopy(config)

        # Check if `_keep_in_fp32_modules` is not None
        # use_keep_in_fp32_modules = cls._keep_in_fp32_modules is not None and (
        #     hf_quantizer is None or getattr(hf_quantizer, "use_keep_in_fp32_modules", False)
        # )
        use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and (mindspore_dtype == ms.float16)

        if use_keep_in_fp32_modules:
            keep_in_fp32_modules = cls._keep_in_fp32_modules
            if not isinstance(keep_in_fp32_modules, list):
                keep_in_fp32_modules = [keep_in_fp32_modules]
        else:
            keep_in_fp32_modules = []

        is_sharded = False
        resolved_model_file = None

        # Determine if we're loading from a directory of sharded checkpoints.
        sharded_metadata = None
        index_file = None
        is_local = os.path.isdir(pretrained_model_name_or_path)
        index_file_kwargs = {
            "is_local": is_local,
            "pretrained_model_name_or_path": pretrained_model_name_or_path,
            "subfolder": subfolder or "",
            "use_safetensors": use_safetensors,
            "cache_dir": cache_dir,
            "variant": variant,
            "force_download": force_download,
            "proxies": proxies,
            "local_files_only": local_files_only,
            "token": token,
            "revision": revision,
            "user_agent": user_agent,
            "commit_hash": commit_hash,
            "dduf_entries": dduf_entries,
        }
        index_file = _fetch_index_file(**index_file_kwargs)
        # In case the index file was not found we still have to consider the legacy format.
        # this becomes applicable when the variant is not None.
        if variant is not None and (index_file is None or not os.path.exists(index_file)):
            index_file = _fetch_index_file_legacy(**index_file_kwargs)
        if index_file is not None and (dduf_entries or index_file.is_file()):
            is_sharded = True

        # load model
        if from_flax:
            raise NotImplementedError("loading flax checkpoint in mindspore model is not yet supported.")
        else:
            # in the case it is sharded, we have already the index
            if is_sharded:
                resolved_model_file, sharded_metadata = _get_checkpoint_shard_files(
                    pretrained_model_name_or_path,
                    index_file,
                    cache_dir=cache_dir,
                    proxies=proxies,
                    local_files_only=local_files_only,
                    token=token,
                    user_agent=user_agent,
                    revision=revision,
                    subfolder=subfolder or "",
                    dduf_entries=dduf_entries,
                )
            elif use_safetensors:
                try:
                    resolved_model_file = _get_model_file(
                        pretrained_model_name_or_path,
                        weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant),
                        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,
                        commit_hash=commit_hash,
                        dduf_entries=dduf_entries,
                    )

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

            if resolved_model_file is None and not is_sharded:
                resolved_model_file = _get_model_file(
                    pretrained_model_name_or_path,
                    weights_name=_add_variant(WEIGHTS_NAME, variant),
                    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,
                    commit_hash=commit_hash,
                    dduf_entries=dduf_entries,
                )

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

        # set dtype to instantiate the model under:
        # 1. If mindspore_dtype is not None, we use that dtype
        # 2. If mindspore_dtype is float8, we don't use _set_default_mindspore_dtype and we downcast after loading the model
        dtype_orig = None  # noqa
        if mindspore_dtype is not None:
            if not isinstance(mindspore_dtype, ms.Type):
                raise ValueError(
                    f"{mindspore_dtype} needs to be of type `mindspore.Type`, e.g. `mindspore.float16`, but is {type(mindspore_dtype)}."
                )

        with no_init_parameters():
            model = cls.from_config(config, **unused_kwargs)

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

        if is_sharded:
            loaded_keys = sharded_metadata["all_checkpoint_keys"]
        else:
            state_dict = _convert_state_dict(model, state_dict)
            loaded_keys = list(state_dict.keys())

        (
            model,
            missing_keys,
            unexpected_keys,
            mismatched_keys,
            offload_index,
            error_msgs,
        ) = cls._load_pretrained_model(
            model,
            state_dict,
            resolved_model_file,
            pretrained_model_name_or_path,
            loaded_keys,
            ignore_mismatched_sizes=ignore_mismatched_sizes,
            dtype=mindspore_dtype,
            keep_in_fp32_modules=keep_in_fp32_modules,
            dduf_entries=dduf_entries,
        )
        loading_info = {
            "missing_keys": missing_keys,
            "unexpected_keys": unexpected_keys,
            "mismatched_keys": mismatched_keys,
            "error_msgs": error_msgs,
        }

        if mindspore_dtype is not None and not use_keep_in_fp32_modules:
            model = model.to(mindspore_dtype)

        model.register_to_config(_name_or_path=pretrained_model_name_or_path)

        # Set model in evaluation mode to deactivate DropOut modules by default
        model.set_train(False)

        if output_loading_info:
            return model, loading_info

        return model

    def to(self, dtype: Optional[ms.Type] = None):
        for p in self.get_parameters():
            p.set_dtype(dtype)
        return self

    def half(self):
        for p in self.get_parameters():
            p.set_dtype(ms.float16)
        return self

    def float(self):
        for p in self.get_parameters():
            p.set_dtype(ms.float32)
        return self

    @classmethod
    def _load_pretrained_model(
        cls,
        model,
        state_dict: OrderedDict,
        resolved_model_file: List[str],
        pretrained_model_name_or_path: Union[str, os.PathLike],
        loaded_keys: List[str],
        ignore_mismatched_sizes: bool = False,
        dtype: Optional[Union[str, ms.Type]] = None,
        keep_in_fp32_modules: Optional[List[str]] = None,
        dduf_entries: Optional[Dict[str, DDUFEntry]] = None,
    ):
        model_state_dict = {k: v for k, v in model.parameters_and_names()}
        expected_keys = list(model_state_dict.keys())
        missing_keys = list(set(expected_keys) - set(loaded_keys))
        unexpected_keys = list(set(loaded_keys) - set(expected_keys))
        # Some models may have keys that are not in the state by design, removing them before needlessly warning
        # the user.
        if cls._keys_to_ignore_on_load_unexpected is not None:
            for pat in cls._keys_to_ignore_on_load_unexpected:
                unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]

        mismatched_keys = []

        error_msgs = []

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

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

        for shard_file in resolved_model_file:
            state_dict = load_state_dict(shard_file, dduf_entries=dduf_entries)

            def _find_mismatched_keys(
                state_dict,
                model_state_dict,
                loaded_keys,
                ignore_mismatched_sizes,
            ):
                mismatched_keys = []
                if ignore_mismatched_sizes:
                    for checkpoint_key in loaded_keys:
                        model_key = checkpoint_key
                        # If the checkpoint is sharded, we may not have the key here.
                        if checkpoint_key not in state_dict:
                            continue

                        if (
                            model_key in model_state_dict
                            and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape
                        ):
                            mismatched_keys.append(
                                (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
                            )
                            del state_dict[checkpoint_key]
                return mismatched_keys

            mismatched_keys += _find_mismatched_keys(
                state_dict,
                model_state_dict,
                loaded_keys,
                ignore_mismatched_sizes,
            )
            if len(resolved_model_file) > 1:
                error_msgs += _load_state_dict_into_model(model, state_dict, is_sharded=True)
            else:
                error_msgs += _load_state_dict_into_model(model, state_dict, is_sharded=False)

        offload_index = None

        if len(error_msgs) > 0:
            error_msg = "\n\t".join(error_msgs)
            if "size mismatch" in error_msg:
                error_msg += (
                    "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
                )
            raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")

        if len(unexpected_keys) > 0:
            logger.warning(
                f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"  # noqa
            )
        else:
            logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")

        if len(missing_keys) > 0:
            logger.warning(
                f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
                f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"
                " TRAIN this model on a down-stream task to be able to use it for predictions and inference."
            )
        elif len(mismatched_keys) == 0:
            logger.info(
                f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"
                f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the"
                f" checkpoint was trained on, you can already use {model.__class__.__name__} for predictions"
                " without further training."
            )
        if len(mismatched_keys) > 0:
            mismatched_warning = "\n".join(
                [
                    f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated"
                    for key, shape1, shape2 in mismatched_keys
                ]
            )
            logger.warning(
                f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
                f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not"
                f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be"
                " able to use it for predictions and inference."
            )

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

    @classmethod
    def _get_signature_keys(cls, obj):
        parameters = inspect.signature(obj.__init__).parameters
        required_parameters = {k: v for k, v in parameters.items() if v.default == inspect._empty}
        optional_parameters = set({k for k, v in parameters.items() if v.default != inspect._empty})
        expected_modules = set(required_parameters.keys()) - {"self"}

        return expected_modules, optional_parameters

    # Adapted from `transformers` modeling_utils.py
    def _get_no_split_modules(self, device_map: str):
        """
        Get the modules of the model that should not be split when using device_map. We iterate through the modules to
        get the underlying `_no_split_modules`.

        Args:
            device_map (`str`):
                The device map value. Options are ["auto", "balanced", "balanced_low_0", "sequential"]

        Returns:
            `List[str]`: List of modules that should not be split
        """
        _no_split_modules = set()
        modules_to_check = [self]
        while len(modules_to_check) > 0:
            module = modules_to_check.pop(-1)
            # if the module does not appear in _no_split_modules, we also check the children
            if module.__class__.__name__ not in _no_split_modules:
                if isinstance(module, ModelMixin):
                    if module._no_split_modules is None:
                        raise ValueError(
                            f"{module.__class__.__name__} does not support `device_map='{device_map}'`. To implement support, the model "
                            "class needs to implement the `_no_split_modules` attribute."
                        )
                    else:
                        _no_split_modules = _no_split_modules | set(module._no_split_modules)
                modules_to_check += list(module.cells())
        return list(_no_split_modules)

    @classmethod
    def _set_default_mindspore_dtype(cls, dtype: ms.Type) -> ms.Type:
        """
        Change the default dtype and return the previous one. This is needed when wanting to instantiate the model
        under specific dtype.

        Args:
            dtype (`mindspore.Type`):
                a floating dtype to set to.

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

        Note `set_default_dtype` currently only works with floating-point types and asserts if for example,
        `ms.int64` is passed. So if a non-float `dtype` is passed this functions will throw an exception.
        """
        raise NotImplementedError("`_set_default_mindspore_dtype` is not yet supported.")

    @property
    def dtype(self) -> ms.Type:
        """
        `mindspore.Type`: The dtype of the module (assuming that all the module parameters have the same dtype).
        """
        return get_parameter_dtype(self)

    def num_parameters(self, only_trainable: bool = False, exclude_embeddings: bool = False) -> int:
        """
        Get number of (trainable or non-embedding) parameters in the module.

        Args:
            only_trainable (`bool`, *optional*, defaults to `False`):
                Whether or not to return only the number of trainable parameters.
            exclude_embeddings (`bool`, *optional*, defaults to `False`):
                Whether or not to return only the number of non-embedding parameters.

        Returns:
            `int`: The number of parameters.

        Example:

        ```py
        from mindone.diffusers import UNet2DConditionModel

        model_id = "runwayml/stable-diffusion-v1-5"
        unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet")
        unet.num_parameters(only_trainable=True)
        859520964
        ```
        """

        if exclude_embeddings:
            embedding_param_names = [
                f"{name}.weight"
                for name, module_type in self.cells_and_names()
                if isinstance(module_type, mint.nn.Embedding)
            ]
            total_parameters = [
                parameter for name, parameter in self.parameters_and_names() if name not in embedding_param_names
            ]
        else:
            total_parameters = list(self.get_parameters())

        total_numel = []

        for param in total_parameters:
            if param.requires_grad or not only_trainable:
                # For 4bit models, we need to multiply the number of parameters by 2 as half of the parameters are
                # used for the 4bit quantization (uint8 tensors are stored)
                total_numel.append(param.numel())

        return sum(total_numel)

    def _set_gradient_checkpointing(self, enable: bool = True) -> None:
        is_gradient_checkpointing_set = False

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

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

    def _fix_state_dict_keys_on_load(self, state_dict: OrderedDict) -> None:
        """
        This function fix the state dict of the model to take into account some changes that were made in the model
        architecture:
        - deprecated attention blocks (happened before we introduced sharded checkpoint,
        so this is why we apply this method only when loading non sharded checkpoints for now)
        """
        deprecated_attention_block_paths = []

        def recursive_find_attn_block(name, module):
            if hasattr(module, "_from_deprecated_attn_block") and module._from_deprecated_attn_block:
                deprecated_attention_block_paths.append(name)

            for sub_name, sub_module in module.name_cells().items():
                sub_name = sub_name if name == "" else f"{name}.{sub_name}"
                recursive_find_attn_block(sub_name, sub_module)

        recursive_find_attn_block("", self)

        # NOTE: we have to check if the deprecated parameters are in the state dict
        # because it is possible we are loading from a state dict that was already
        # converted

        for path in deprecated_attention_block_paths:
            # group_norm path stays the same

            # query -> to_q
            if f"{path}.query.weight" in state_dict:
                state_dict[f"{path}.to_q.weight"] = state_dict.pop(f"{path}.query.weight")
            if f"{path}.query.bias" in state_dict:
                state_dict[f"{path}.to_q.bias"] = state_dict.pop(f"{path}.query.bias")

            # key -> to_k
            if f"{path}.key.weight" in state_dict:
                state_dict[f"{path}.to_k.weight"] = state_dict.pop(f"{path}.key.weight")
            if f"{path}.key.bias" in state_dict:
                state_dict[f"{path}.to_k.bias"] = state_dict.pop(f"{path}.key.bias")

            # value -> to_v
            if f"{path}.value.weight" in state_dict:
                state_dict[f"{path}.to_v.weight"] = state_dict.pop(f"{path}.value.weight")
            if f"{path}.value.bias" in state_dict:
                state_dict[f"{path}.to_v.bias"] = state_dict.pop(f"{path}.value.bias")

            # proj_attn -> to_out.0
            if f"{path}.proj_attn.weight" in state_dict:
                state_dict[f"{path}.to_out.0.weight"] = state_dict.pop(f"{path}.proj_attn.weight")
            if f"{path}.proj_attn.bias" in state_dict:
                state_dict[f"{path}.to_out.0.bias"] = state_dict.pop(f"{path}.proj_attn.bias")

        # TODO : MindSpore 2.6 share weight bug. Unable to load WTE and LM-Head layer weights properly. It will be
        #  deleted until fixed load_state_dict_into_model and parameters_and_namesใ€‚
        if hasattr(self, "wte_lm_share") and self.wte_lm_share:
            state_dict["transformer.transformer.wte.embedding_table"] = state_dict["transformer.lm_head.weight"]

        return state_dict

    def get_submodule(self, target: str) -> nn.Cell:
        """Return the submodule given by ``target`` if it exists, otherwise throw an error.

        For example, let's say you have an ``nn.Cell`` ``A`` that
        looks like this:

        .. code-block:: text

            A(
                (net_b): Module(
                    (net_c): Module(
                        (conv): Conv2d(16, 33, kernel_size=(3, 3), stride=(2, 2))
                    )
                    (linear): Dense(input_channels=100, output_channels=200, has_bias=True)
                )
            )

        (The diagram shows an ``nn.Cell`` ``A``. ``A`` has a nested
        submodule ``net_b``, which itself has two submodules ``net_c``
        and ``linear``. ``net_c`` then has a submodule ``conv``.)

        To check whether or not we have the ``linear`` submodule, we
        would call ``get_submodule("net_b.linear")``. To check whether
        we have the ``conv`` submodule, we would call
        ``get_submodule("net_b.net_c.conv")``.

        The runtime of ``get_submodule`` is bounded by the degree
        of module nesting in ``target``. A query against
        ``named_modules`` achieves the same result, but it is O(N) in
        the number of transitive modules. So, for a simple check to see
        if some submodule exists, ``get_submodule`` should always be
        used.

        Args:
            target: The fully-qualified string name of the submodule
                to look for. (See above example for how to specify a
                fully-qualified string.)

        Returns:
            nn.Cell: The submodule referenced by ``target``

        Raises:
            AttributeError: If the target string references an invalid
                path or resolves to something that is not an
                ``nn.Cell``
        """
        if target == "":
            return self

        atoms: List[str] = target.split(".")
        mod: nn.Cell = self

        for item in atoms:
            if not hasattr(mod, item):
                raise AttributeError(mod.cls_name + " has no " "attribute `" + item + "`")

            mod = getattr(mod, item)

            if not isinstance(mod, nn.Cell):
                raise AttributeError("`" + item + "` is not " "an nn.Module")

        return mod

mindone.diffusers.ModelMixin.dtype: ms.Type property

mindspore.Type: The dtype of the module (assuming that all the module parameters have the same dtype).

mindone.diffusers.ModelMixin.is_gradient_checkpointing: bool property

Whether gradient checkpointing is activated for this model or not.

mindone.diffusers.ModelMixin.__getattr__(name)

The only reason we overwrite getattr here is to gracefully deprecate accessing config attributes directly. See https://github.com/huggingface/diffusers/pull/3129 We need to overwrite getattr here in addition so that we don't trigger nn.Cell's getattr': https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module

Source code in mindone/diffusers/models/modeling_utils.py
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
def __getattr__(self, name: str) -> Any:
    """The only reason we overwrite `getattr` here is to gracefully deprecate accessing
    config attributes directly. See https://github.com/huggingface/diffusers/pull/3129 We need to overwrite
    __getattr__ here in addition so that we don't trigger `nn.Cell`'s __getattr__':
    https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
    """

    is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name)
    is_attribute = name in self.__dict__

    if is_in_config and not is_attribute:
        deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'unet.config.{name}'."  # noqa: E501
        deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False, stacklevel=3)
        return self._internal_dict[name]

    # call PyTorch's https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
    return super().__getattr__(name)

mindone.diffusers.ModelMixin.disable_gradient_checkpointing()

Deactivates gradient checkpointing for the current model (may be referred to as activation checkpointing or checkpoint activations in other frameworks).

Source code in mindone/diffusers/models/modeling_utils.py
221
222
223
224
225
226
227
def disable_gradient_checkpointing(self) -> None:
    """
    Deactivates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
    *checkpoint activations* in other frameworks).
    """
    if self._supports_gradient_checkpointing:
        self._set_gradient_checkpointing(enable=False)

mindone.diffusers.ModelMixin.disable_xformers_memory_efficient_attention()

Disable memory efficient attention from xFormers.

Source code in mindone/diffusers/models/modeling_utils.py
326
327
328
329
330
def disable_xformers_memory_efficient_attention(self) -> None:
    r"""
    Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
    """
    self.set_use_memory_efficient_attention_xformers(False)

mindone.diffusers.ModelMixin.enable_flash_sdp(enabled)

.. warning:: This flag is beta and subject to change.

Enables or disables flash scaled dot product attention.

Source code in mindone/diffusers/models/modeling_utils.py
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
def enable_flash_sdp(self, enabled: bool):
    r"""
    .. warning:: This flag is beta and subject to change.

    Enables or disables flash scaled dot product attention.
    """

    # Recursively walk through all the children.
    # Any children which exposes the enable_flash_sdp method
    # gets the message
    def fn_recursive_set_mem_eff(module: nn.Cell):
        if hasattr(module, "enable_flash_sdp"):
            module.enable_flash_sdp(enabled)

        for child in module.cells():
            fn_recursive_set_mem_eff(child)

    for module in self.cells():
        if isinstance(module, nn.Cell):
            fn_recursive_set_mem_eff(module)

mindone.diffusers.ModelMixin.enable_gradient_checkpointing(gradient_checkpointing_func=None)

Activates gradient checkpointing for the current model (may be referred to as activation checkpointing or checkpoint activations in other frameworks).

PARAMETER DESCRIPTION
gradient_checkpointing_func

The function to use for gradient checkpointing. If None, the default MindSpore checkpointing function is used (mindspore.nn.Cell.recompute_).

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

Source code in mindone/diffusers/models/modeling_utils.py
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
def enable_gradient_checkpointing(self, gradient_checkpointing_func: Optional[Callable] = None) -> None:
    """
    Activates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
    *checkpoint activations* in other frameworks).

    Args:
        gradient_checkpointing_func (`Callable`, *optional*):
            The function to use for gradient checkpointing. If `None`, the default MindSpore checkpointing function
            is used (`mindspore.nn.Cell.recompute_`).
    """
    if not self._supports_gradient_checkpointing:
        raise ValueError(
            f"{self.__class__.__name__} does not support gradient checkpointing. Please make sure to set the boolean attribute "
            f"`_supports_gradient_checkpointing` to `True` in the class definition."
        )

    if gradient_checkpointing_func is None:

        def _gradient_checkpointing_func(module, *args):
            module.recompute_(mode=True)
            return module

        gradient_checkpointing_func = _gradient_checkpointing_func

    self._set_gradient_checkpointing(enable=True)

mindone.diffusers.ModelMixin.enable_group_offload(onload_device='Ascend', offload_device='CPU', offload_type='block_level', num_blocks_per_group=None, non_blocking=False, use_stream=False, record_stream=False, low_cpu_mem_usage=False)

Activates group offloading for the current model.

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

Example:

```python
>>> from mindone.diffusers import CogVideoXTransformer3DModel

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

>>> transformer.enable_group_offload(
...     onload_device="Ascend",
...     offload_device="CPU",
...     offload_type="leaf_level",
...     use_stream=True,
... )
```
Source code in mindone/diffusers/models/modeling_utils.py
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
def enable_group_offload(
    self,
    onload_device: str = "Ascend",
    offload_device: str = "CPU",
    offload_type: str = "block_level",
    num_blocks_per_group: Optional[int] = None,
    non_blocking: bool = False,
    use_stream: bool = False,
    record_stream: bool = False,
    low_cpu_mem_usage=False,
) -> None:
    r"""
    Activates group offloading for the current model.

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

    Example:

        ```python
        >>> from mindone.diffusers import CogVideoXTransformer3DModel

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

        >>> transformer.enable_group_offload(
        ...     onload_device="Ascend",
        ...     offload_device="CPU",
        ...     offload_type="leaf_level",
        ...     use_stream=True,
        ... )
        ```
    """
    raise NotImplementedError("`enable_group_offload` is not yet supported.")

mindone.diffusers.ModelMixin.enable_layerwise_casting(storage_dtype, compute_dtype=None, skip_modules_pattern=None, skip_modules_classes=None, non_blocking=False)

Activates layerwise casting for the current model.

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

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

Example

Using [~models.ModelMixin.enable_layerwise_casting]:

>>> from mindone.diffusers import CogVideoXTransformer3DModel

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

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

The dtype to which the model should be cast for storage.

TYPE: `mindspore.Type`

compute_dtype

The dtype to which the model weights should be cast during the forward pass.

TYPE: `mindspore.Type` DEFAULT: None

skip_modules_pattern

A list of patterns to match the names of the modules to skip during the layerwise casting process. If set to None, default skip patterns are used to ignore certain internal layers of modules and PEFT layers.

TYPE: `Tuple[str, ...]`, *optional* DEFAULT: None

skip_modules_classes

A list of module classes to skip during the layerwise casting process.

TYPE: `Tuple[Type[nn.Cell], ...]`, *optional* DEFAULT: None

non_blocking

If True, the weight casting operations are non-blocking.

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

Source code in mindone/diffusers/models/modeling_utils.py
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
def enable_layerwise_casting(
    self,
    storage_dtype: ms.Type,
    compute_dtype: Optional[ms.Type] = None,
    skip_modules_pattern: Optional[Tuple[str, ...]] = None,
    skip_modules_classes: Optional[Tuple[Type[nn.Cell], ...]] = None,
    non_blocking: bool = False,
) -> None:
    r"""
    Activates layerwise casting for the current model.

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

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

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

        ```python
        >>> from mindone.diffusers import CogVideoXTransformer3DModel

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

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

    Args:
        storage_dtype (`mindspore.Type`):
            The dtype to which the model should be cast for storage.
        compute_dtype (`mindspore.Type`):
            The dtype to which the model weights should be cast during the forward pass.
        skip_modules_pattern (`Tuple[str, ...]`, *optional*):
            A list of patterns to match the names of the modules to skip during the layerwise casting process. If
            set to `None`, default skip patterns are used to ignore certain internal layers of modules and PEFT
            layers.
        skip_modules_classes (`Tuple[Type[nn.Cell], ...]`, *optional*):
            A list of module classes to skip during the layerwise casting process.
        non_blocking (`bool`, *optional*, defaults to `False`):
            If `True`, the weight casting operations are non-blocking.
    """
    raise NotImplementedError("`enable_layerwise_casting` is not yet supported.")

mindone.diffusers.ModelMixin.enable_xformers_memory_efficient_attention(attention_op=None)

Enable memory efficient attention from xFormers.

When this option is enabled, you should observe lower GPU memory usage and a potential speed up during inference. Speed up during training is not guaranteed.

โš ๏ธ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes precedent.

PARAMETER DESCRIPTION
attention_op

Not supported for now.

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

>>> import mindspore as ms
>>> from mindone.diffusers import UNet2DConditionModel

>>> model = UNet2DConditionModel.from_pretrained(
...     "stabilityai/stable-diffusion-2-1", subfolder="unet", mindspore_dtype=ms.float16
... )
>>> model.enable_xformers_memory_efficient_attention()
Source code in mindone/diffusers/models/modeling_utils.py
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
def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None) -> None:
    r"""
    Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).

    When this option is enabled, you should observe lower GPU memory usage and a potential speed up during
    inference. Speed up during training is not guaranteed.

    <Tip warning={true}>

    โš ๏ธ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes
    precedent.

    </Tip>

    Parameters:
        attention_op (`Callable`, *optional*):
            Not supported for now.

    Examples:

    ```py
    >>> import mindspore as ms
    >>> from mindone.diffusers import UNet2DConditionModel

    >>> model = UNet2DConditionModel.from_pretrained(
    ...     "stabilityai/stable-diffusion-2-1", subfolder="unet", mindspore_dtype=ms.float16
    ... )
    >>> model.enable_xformers_memory_efficient_attention()
    ```
    """
    self.set_use_memory_efficient_attention_xformers(True, attention_op)

mindone.diffusers.ModelMixin.from_pretrained(pretrained_model_name_or_path, **kwargs) classmethod

Instantiate a pretrained PyTorch model from a pretrained model configuration.

The model is set in evaluation mode - model.eval() - by default, and dropout modules are deactivated. To train the model, set it back in training mode with model.train().

PARAMETER DESCRIPTION
pretrained_model_name_or_path

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`].

TYPE: `str` or `os.PathLike`, *optional*

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*

mindspore_dtype

Override the default mindspore.Type and load the model with another dtype. If "auto" is passed, the dtype is automatically derived from the model's weights.

TYPE: `str` or `mindspore.Type`, *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*

output_loading_info

Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.

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

local_files_only(`bool`,

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: *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"`

from_flax

Load the model weights from a Flax checkpoint save file.

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

subfolder

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

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

mirror

Mirror source to resolve accessibility issues if you're downloading a model in China. We do not guarantee the timeliness or safety of the source, and you should refer to the mirror site for more information.

TYPE: `str`, *optional*

variant

Load weights from a specified variant filename such as "fp16" or "ema". This is ignored when loading from_flax.

TYPE: `str`, *optional*

use_safetensors

If set to None, the safetensors weights are downloaded if they're available and if the safetensors library is installed. If set to True, the model is forcibly loaded from safetensors weights. If set to False, safetensors weights are not loaded.

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

disable_mmap

Whether to disable mmap when loading a Safetensors model. This option can perform better when the model is on a network mount or hard drive, which may not handle the seeky-ness of mmap very well.

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

To use private or gated models, log-in with huggingface-cli login. You can also activate the special "offline-mode" to use this method in a firewalled environment.

Example:

from mindone.diffusers import UNet2DConditionModel

unet = UNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="unet")

If you get the error message below, you need to finetune the weights for your downstream task:

Some weights of UNet2DConditionModel were not initialized from the model checkpoint at
runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
- conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
Source code in mindone/diffusers/models/modeling_utils.py
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
@classmethod
@validate_hf_hub_args
def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs) -> Self:
    r"""
    Instantiate a pretrained PyTorch model from a pretrained model configuration.

    The model is set in evaluation mode - `model.eval()` - by default, and dropout modules are deactivated. To
    train the model, set it back in training mode with `model.train()`.

    Parameters:
        pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
            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`].

        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.
        mindspore_dtype (`str` or `mindspore.Type`, *optional*):
            Override the default `mindspore.Type` and load the model with another dtype. If `"auto"` is passed, the
            dtype is automatically derived from the model's weights.
        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.
        output_loading_info (`bool`, *optional*, defaults to `False`):
            Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
        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.
        from_flax (`bool`, *optional*, defaults to `False`):
            Load the model weights from a Flax checkpoint save file.
        subfolder (`str`, *optional*, defaults to `""`):
            The subfolder location of a model file within a larger model repository on the Hub or locally.
        mirror (`str`, *optional*):
            Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
            guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
            information.
        variant (`str`, *optional*):
            Load weights from a specified `variant` filename such as `"fp16"` or `"ema"`. This is ignored when
            loading `from_flax`.
        use_safetensors (`bool`, *optional*, defaults to `None`):
            If set to `None`, the `safetensors` weights are downloaded if they're available **and** if the
            `safetensors` library is installed. If set to `True`, the model is forcibly loaded from `safetensors`
            weights. If set to `False`, `safetensors` weights are not loaded.
        disable_mmap ('bool', *optional*, defaults to 'False'):
            Whether to disable mmap when loading a Safetensors model. This option can perform better when the model
            is on a network mount or hard drive, which may not handle the seeky-ness of mmap very well.

    <Tip>

    To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with
    `huggingface-cli login`. You can also activate the special
    ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use this method in a
    firewalled environment.

    </Tip>

    Example:

    ```py
    from mindone.diffusers import UNet2DConditionModel

    unet = UNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="unet")
    ```

    If you get the error message below, you need to finetune the weights for your downstream task:

    ```bash
    Some weights of UNet2DConditionModel were not initialized from the model checkpoint at
    runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
    - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
    You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
    ```
    """
    cache_dir = kwargs.pop("cache_dir", None)
    ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
    force_download = kwargs.pop("force_download", False)
    from_flax = kwargs.pop("from_flax", False)
    proxies = kwargs.pop("proxies", None)
    output_loading_info = kwargs.pop("output_loading_info", False)
    local_files_only = kwargs.pop("local_files_only", None)
    token = kwargs.pop("token", None)
    revision = kwargs.pop("revision", None)
    mindspore_dtype = kwargs.pop("mindspore_dtype", None)
    subfolder = kwargs.pop("subfolder", None)
    variant = kwargs.pop("variant", None)
    use_safetensors = kwargs.pop("use_safetensors", None)
    dduf_entries: Optional[Dict[str, DDUFEntry]] = kwargs.pop("dduf_entries", None)
    disable_mmap = kwargs.pop("disable_mmap", False)

    if mindspore_dtype is not None and not isinstance(mindspore_dtype, ms.Type):
        mindspore_dtype = ms.float32
        logger.warning(
            f"Passed `mindspore_dtype` {mindspore_dtype} is not a `ms.Type`. Defaulting to `ms.float32`."
        )

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

    user_agent = {
        "diffusers": __version__,
        "file_type": "model",
        "framework": "pytorch",
    }
    unused_kwargs = {}

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

    # load config
    config, unused_kwargs, commit_hash = cls.load_config(
        config_path,
        cache_dir=cache_dir,
        return_unused_kwargs=True,
        return_commit_hash=True,
        force_download=force_download,
        proxies=proxies,
        local_files_only=local_files_only,
        token=token,
        revision=revision,
        subfolder=subfolder,
        user_agent=user_agent,
        dduf_entries=dduf_entries,
        **kwargs,
    )
    # no in-place modification of the original config.
    config = copy.deepcopy(config)

    # Check if `_keep_in_fp32_modules` is not None
    # use_keep_in_fp32_modules = cls._keep_in_fp32_modules is not None and (
    #     hf_quantizer is None or getattr(hf_quantizer, "use_keep_in_fp32_modules", False)
    # )
    use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and (mindspore_dtype == ms.float16)

    if use_keep_in_fp32_modules:
        keep_in_fp32_modules = cls._keep_in_fp32_modules
        if not isinstance(keep_in_fp32_modules, list):
            keep_in_fp32_modules = [keep_in_fp32_modules]
    else:
        keep_in_fp32_modules = []

    is_sharded = False
    resolved_model_file = None

    # Determine if we're loading from a directory of sharded checkpoints.
    sharded_metadata = None
    index_file = None
    is_local = os.path.isdir(pretrained_model_name_or_path)
    index_file_kwargs = {
        "is_local": is_local,
        "pretrained_model_name_or_path": pretrained_model_name_or_path,
        "subfolder": subfolder or "",
        "use_safetensors": use_safetensors,
        "cache_dir": cache_dir,
        "variant": variant,
        "force_download": force_download,
        "proxies": proxies,
        "local_files_only": local_files_only,
        "token": token,
        "revision": revision,
        "user_agent": user_agent,
        "commit_hash": commit_hash,
        "dduf_entries": dduf_entries,
    }
    index_file = _fetch_index_file(**index_file_kwargs)
    # In case the index file was not found we still have to consider the legacy format.
    # this becomes applicable when the variant is not None.
    if variant is not None and (index_file is None or not os.path.exists(index_file)):
        index_file = _fetch_index_file_legacy(**index_file_kwargs)
    if index_file is not None and (dduf_entries or index_file.is_file()):
        is_sharded = True

    # load model
    if from_flax:
        raise NotImplementedError("loading flax checkpoint in mindspore model is not yet supported.")
    else:
        # in the case it is sharded, we have already the index
        if is_sharded:
            resolved_model_file, sharded_metadata = _get_checkpoint_shard_files(
                pretrained_model_name_or_path,
                index_file,
                cache_dir=cache_dir,
                proxies=proxies,
                local_files_only=local_files_only,
                token=token,
                user_agent=user_agent,
                revision=revision,
                subfolder=subfolder or "",
                dduf_entries=dduf_entries,
            )
        elif use_safetensors:
            try:
                resolved_model_file = _get_model_file(
                    pretrained_model_name_or_path,
                    weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant),
                    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,
                    commit_hash=commit_hash,
                    dduf_entries=dduf_entries,
                )

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

        if resolved_model_file is None and not is_sharded:
            resolved_model_file = _get_model_file(
                pretrained_model_name_or_path,
                weights_name=_add_variant(WEIGHTS_NAME, variant),
                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,
                commit_hash=commit_hash,
                dduf_entries=dduf_entries,
            )

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

    # set dtype to instantiate the model under:
    # 1. If mindspore_dtype is not None, we use that dtype
    # 2. If mindspore_dtype is float8, we don't use _set_default_mindspore_dtype and we downcast after loading the model
    dtype_orig = None  # noqa
    if mindspore_dtype is not None:
        if not isinstance(mindspore_dtype, ms.Type):
            raise ValueError(
                f"{mindspore_dtype} needs to be of type `mindspore.Type`, e.g. `mindspore.float16`, but is {type(mindspore_dtype)}."
            )

    with no_init_parameters():
        model = cls.from_config(config, **unused_kwargs)

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

    if is_sharded:
        loaded_keys = sharded_metadata["all_checkpoint_keys"]
    else:
        state_dict = _convert_state_dict(model, state_dict)
        loaded_keys = list(state_dict.keys())

    (
        model,
        missing_keys,
        unexpected_keys,
        mismatched_keys,
        offload_index,
        error_msgs,
    ) = cls._load_pretrained_model(
        model,
        state_dict,
        resolved_model_file,
        pretrained_model_name_or_path,
        loaded_keys,
        ignore_mismatched_sizes=ignore_mismatched_sizes,
        dtype=mindspore_dtype,
        keep_in_fp32_modules=keep_in_fp32_modules,
        dduf_entries=dduf_entries,
    )
    loading_info = {
        "missing_keys": missing_keys,
        "unexpected_keys": unexpected_keys,
        "mismatched_keys": mismatched_keys,
        "error_msgs": error_msgs,
    }

    if mindspore_dtype is not None and not use_keep_in_fp32_modules:
        model = model.to(mindspore_dtype)

    model.register_to_config(_name_or_path=pretrained_model_name_or_path)

    # Set model in evaluation mode to deactivate DropOut modules by default
    model.set_train(False)

    if output_loading_info:
        return model, loading_info

    return model

mindone.diffusers.ModelMixin.get_submodule(target)

Return the submodule given by target if it exists, otherwise throw an error.

For example, let's say you have an nn.Cell A that looks like this:

.. code-block:: text

A(
    (net_b): Module(
        (net_c): Module(
            (conv): Conv2d(16, 33, kernel_size=(3, 3), stride=(2, 2))
        )
        (linear): Dense(input_channels=100, output_channels=200, has_bias=True)
    )
)

(The diagram shows an nn.Cell A. A has a nested submodule net_b, which itself has two submodules net_c and linear. net_c then has a submodule conv.)

To check whether or not we have the linear submodule, we would call get_submodule("net_b.linear"). To check whether we have the conv submodule, we would call get_submodule("net_b.net_c.conv").

The runtime of get_submodule is bounded by the degree of module nesting in target. A query against named_modules achieves the same result, but it is O(N) in the number of transitive modules. So, for a simple check to see if some submodule exists, get_submodule should always be used.

PARAMETER DESCRIPTION
target

The fully-qualified string name of the submodule to look for. (See above example for how to specify a fully-qualified string.)

TYPE: str

RETURNS DESCRIPTION
Cell

nn.Cell: The submodule referenced by target

RAISES DESCRIPTION
AttributeError

If the target string references an invalid path or resolves to something that is not an nn.Cell

Source code in mindone/diffusers/models/modeling_utils.py
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
def get_submodule(self, target: str) -> nn.Cell:
    """Return the submodule given by ``target`` if it exists, otherwise throw an error.

    For example, let's say you have an ``nn.Cell`` ``A`` that
    looks like this:

    .. code-block:: text

        A(
            (net_b): Module(
                (net_c): Module(
                    (conv): Conv2d(16, 33, kernel_size=(3, 3), stride=(2, 2))
                )
                (linear): Dense(input_channels=100, output_channels=200, has_bias=True)
            )
        )

    (The diagram shows an ``nn.Cell`` ``A``. ``A`` has a nested
    submodule ``net_b``, which itself has two submodules ``net_c``
    and ``linear``. ``net_c`` then has a submodule ``conv``.)

    To check whether or not we have the ``linear`` submodule, we
    would call ``get_submodule("net_b.linear")``. To check whether
    we have the ``conv`` submodule, we would call
    ``get_submodule("net_b.net_c.conv")``.

    The runtime of ``get_submodule`` is bounded by the degree
    of module nesting in ``target``. A query against
    ``named_modules`` achieves the same result, but it is O(N) in
    the number of transitive modules. So, for a simple check to see
    if some submodule exists, ``get_submodule`` should always be
    used.

    Args:
        target: The fully-qualified string name of the submodule
            to look for. (See above example for how to specify a
            fully-qualified string.)

    Returns:
        nn.Cell: The submodule referenced by ``target``

    Raises:
        AttributeError: If the target string references an invalid
            path or resolves to something that is not an
            ``nn.Cell``
    """
    if target == "":
        return self

    atoms: List[str] = target.split(".")
    mod: nn.Cell = self

    for item in atoms:
        if not hasattr(mod, item):
            raise AttributeError(mod.cls_name + " has no " "attribute `" + item + "`")

        mod = getattr(mod, item)

        if not isinstance(mod, nn.Cell):
            raise AttributeError("`" + item + "` is not " "an nn.Module")

    return mod

mindone.diffusers.ModelMixin.num_parameters(only_trainable=False, exclude_embeddings=False)

Get number of (trainable or non-embedding) parameters in the module.

PARAMETER DESCRIPTION
only_trainable

Whether or not to return only the number of trainable parameters.

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

exclude_embeddings

Whether or not to return only the number of non-embedding parameters.

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

RETURNS DESCRIPTION
int

int: The number of parameters.

from mindone.diffusers import UNet2DConditionModel

model_id = "runwayml/stable-diffusion-v1-5"
unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet")
unet.num_parameters(only_trainable=True)
859520964
Source code in mindone/diffusers/models/modeling_utils.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
def num_parameters(self, only_trainable: bool = False, exclude_embeddings: bool = False) -> int:
    """
    Get number of (trainable or non-embedding) parameters in the module.

    Args:
        only_trainable (`bool`, *optional*, defaults to `False`):
            Whether or not to return only the number of trainable parameters.
        exclude_embeddings (`bool`, *optional*, defaults to `False`):
            Whether or not to return only the number of non-embedding parameters.

    Returns:
        `int`: The number of parameters.

    Example:

    ```py
    from mindone.diffusers import UNet2DConditionModel

    model_id = "runwayml/stable-diffusion-v1-5"
    unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet")
    unet.num_parameters(only_trainable=True)
    859520964
    ```
    """

    if exclude_embeddings:
        embedding_param_names = [
            f"{name}.weight"
            for name, module_type in self.cells_and_names()
            if isinstance(module_type, mint.nn.Embedding)
        ]
        total_parameters = [
            parameter for name, parameter in self.parameters_and_names() if name not in embedding_param_names
        ]
    else:
        total_parameters = list(self.get_parameters())

    total_numel = []

    for param in total_parameters:
        if param.requires_grad or not only_trainable:
            # For 4bit models, we need to multiply the number of parameters by 2 as half of the parameters are
            # used for the 4bit quantization (uint8 tensors are stored)
            total_numel.append(param.numel())

    return sum(total_numel)

mindone.diffusers.ModelMixin.save_pretrained(save_directory, is_main_process=True, save_function=None, safe_serialization=True, variant=None, max_shard_size='10GB', push_to_hub=False, **kwargs)

Save a model and its configuration file to a directory so that it can be reloaded using the [~models.ModelMixin.from_pretrained] class method.

PARAMETER DESCRIPTION
save_directory

Directory to save a model and its configuration file to. Will be created if it doesn't exist.

TYPE: `str` or `os.PathLike`

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 PyTorch way with pickle.

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

variant

If specified, weights are saved in the format pytorch_model.<variant>.bin.

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

max_shard_size

The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size lower than this size. If expressed as a string, needs to be digits followed by a unit (like "5GB"). If expressed as an integer, the unit is bytes. Note that this limit will be decreased after a certain period of time (starting from Oct 2024) to allow users to upgrade to the latest version of diffusers. This is to establish a common default size for this argument across different libraries in the Hugging Face ecosystem (transformers, and accelerate, for example).

TYPE: `int` or `str`, defaults to `"10GB"` DEFAULT: '10GB'

push_to_hub

Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the repository you want to push to with repo_id (will default to the name of save_directory in your namespace).

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

kwargs

Additional keyword arguments passed along to the [~utils.PushToHubMixin.push_to_hub] method.

TYPE: `Dict[str, Any]`, *optional* DEFAULT: {}

Source code in mindone/diffusers/models/modeling_utils.py
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
def save_pretrained(
    self,
    save_directory: Union[str, os.PathLike],
    is_main_process: bool = True,
    save_function: Optional[Callable] = None,
    safe_serialization: bool = True,
    variant: Optional[str] = None,
    max_shard_size: Union[int, str] = "10GB",
    push_to_hub: bool = False,
    **kwargs,
):
    """
    Save a model and its configuration file to a directory so that it can be reloaded using the
    [`~models.ModelMixin.from_pretrained`] class method.

    Arguments:
        save_directory (`str` or `os.PathLike`):
            Directory to save a model and its configuration file to. Will be created if it doesn't exist.
        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 PyTorch way with `pickle`.
        variant (`str`, *optional*):
            If specified, weights are saved in the format `pytorch_model.<variant>.bin`.
        max_shard_size (`int` or `str`, defaults to `"10GB"`):
            The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size
            lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5GB"`).
            If expressed as an integer, the unit is bytes. Note that this limit will be decreased after a certain
            period of time (starting from Oct 2024) to allow users to upgrade to the latest version of `diffusers`.
            This is to establish a common default size for this argument across different libraries in the Hugging
            Face ecosystem (`transformers`, and `accelerate`, for example).
        push_to_hub (`bool`, *optional*, defaults to `False`):
            Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the
            repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
            namespace).
        kwargs (`Dict[str, Any]`, *optional*):
            Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
    """
    if os.path.isfile(save_directory):
        logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
        return

    weights_name = SAFETENSORS_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
    weights_name = _add_variant(weights_name, variant)
    weights_name_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(
        ".safetensors", "{suffix}.safetensors"
    )

    os.makedirs(save_directory, exist_ok=True)

    if push_to_hub:
        commit_message = kwargs.pop("commit_message", None)
        private = kwargs.pop("private", None)
        create_pr = kwargs.pop("create_pr", False)
        token = kwargs.pop("token", None)
        repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
        repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id

    # Only save the model itself if we are using distributed training
    model_to_save = self

    # Attach architecture to the config
    # Save the config
    if is_main_process:
        model_to_save.save_config(save_directory)

    # Save the model
    state_dict = {k: v for k, v in model_to_save.parameters_and_names()}

    # Save the model
    state_dict_split = split_torch_state_dict_into_shards(
        state_dict, max_shard_size=max_shard_size, filename_pattern=weights_name_pattern
    )

    # Clean the folder from a previous save
    if is_main_process:
        for filename in os.listdir(save_directory):
            if filename in state_dict_split.filename_to_tensors.keys():
                continue
            full_filename = os.path.join(save_directory, filename)
            if not os.path.isfile(full_filename):
                continue
            weights_without_ext = weights_name_pattern.replace(".bin", "").replace(".safetensors", "")
            weights_without_ext = weights_without_ext.replace("{suffix}", "")
            filename_without_ext = filename.replace(".bin", "").replace(".safetensors", "")
            # make sure that file to be deleted matches format of sharded file, e.g. pytorch_model-00001-of-00005
            if (
                filename.startswith(weights_without_ext)
                and _REGEX_SHARD.fullmatch(filename_without_ext) is not None
            ):
                os.remove(full_filename)

    for filename, tensors in state_dict_split.filename_to_tensors.items():
        shard = {tensor: state_dict[tensor].contiguous() for tensor in tensors}
        filepath = os.path.join(save_directory, filename)
        if safe_serialization:
            # At some point we will need to deal better with save_function (used for TPU and other distributed
            # joyfulness), but for now this enough.
            safe_save_file(shard, filepath, metadata={"format": "np"})
        else:
            ms.save_checkpoint(shard, filepath)

    if state_dict_split.is_sharded:
        index = {
            "metadata": state_dict_split.metadata,
            "weight_map": state_dict_split.tensor_to_filename,
        }
        save_index_file = SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
        save_index_file = os.path.join(save_directory, _add_variant(save_index_file, variant))
        # Save the index as well
        with open(save_index_file, "w", encoding="utf-8") as f:
            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
            f.write(content)
        logger.info(
            f"The model is bigger than the maximum size per checkpoint ({max_shard_size}) and is going to be "
            f"split in {len(state_dict_split.filename_to_tensors)} checkpoint shards. You can find where each parameters has been saved in the "
            f"index located at {save_index_file}."
        )
    else:
        path_to_weights = os.path.join(save_directory, weights_name)
        logger.info(f"Model weights saved in {path_to_weights}")

    if push_to_hub:
        # Create a new empty model card and eventually tag it
        model_card = load_or_create_model_card(repo_id, token=token)
        model_card = populate_model_card(model_card)
        model_card.save(Path(save_directory, "README.md").as_posix())

        self._upload_folder(
            save_directory,
            repo_id,
            token=token,
            commit_message=commit_message,
            create_pr=create_pr,
        )

mindone.diffusers.ModelMixin.set_flash_attention_force_cast_dtype(force_cast_dtype)

Since the flash-attention operator in MindSpore only supports float16 and bfloat16 data types, we need to manually set whether to force data type conversion.

When the attention interface encounters data of an unsupported data type, if force_cast_dtype is not None, the function will forcibly convert the data to force_cast_dtype for computation and then restore it to the original data type afterward. If force_cast_dtype is None, it will fall back to the original attention calculation using mathematical formulas.

PARAMETER DESCRIPTION
force_cast_dtype

The data type to which the input data should be forcibly converted. If None, no forced

TYPE: Optional

Source code in mindone/diffusers/models/modeling_utils.py
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
def set_flash_attention_force_cast_dtype(self, force_cast_dtype: Optional[ms.Type]):
    r"""
    Since the flash-attention operator in MindSpore only supports float16 and bfloat16 data types, we need to manually
    set whether to force data type conversion.

    When the attention interface encounters data of an unsupported data type,
    if `force_cast_dtype` is not None, the function will forcibly convert the data to `force_cast_dtype` for computation
    and then restore it to the original data type afterward. If `force_cast_dtype` is None, it will fall back to the
    original attention calculation using mathematical formulas.

    Parameters:
        force_cast_dtype (Optional): The data type to which the input data should be forcibly converted. If None, no forced
        conversion is performed.
    """

    # Recursively walk through all the children.
    # Any children which exposes the set_flash_attention_force_cast_dtype method
    # gets the message
    def fn_recursive_set_mem_eff(module: nn.Cell):
        if hasattr(module, "set_flash_attention_force_cast_dtype"):
            module.set_flash_attention_force_cast_dtype(force_cast_dtype)

        for child in module.cells():
            fn_recursive_set_mem_eff(child)

    for module in self.cells():
        if isinstance(module, nn.Cell):
            fn_recursive_set_mem_eff(module)

mindone.diffusers.utils.PushToHubMixin

A Mixin to push a model, scheduler, or pipeline to the Hugging Face Hub.

Source code in mindone/diffusers/utils/hub_utils.py
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
class PushToHubMixin:
    """
    A Mixin to push a model, scheduler, or pipeline to the Hugging Face Hub.
    """

    def _upload_folder(
        self,
        working_dir: Union[str, os.PathLike],
        repo_id: str,
        token: Optional[str] = None,
        commit_message: Optional[str] = None,
        create_pr: bool = False,
    ):
        """
        Uploads all files in `working_dir` to `repo_id`.
        """
        if commit_message is None:
            if "Model" in self.__class__.__name__:
                commit_message = "Upload model"
            elif "Scheduler" in self.__class__.__name__:
                commit_message = "Upload scheduler"
            else:
                commit_message = f"Upload {self.__class__.__name__}"

        logger.info(f"Uploading the files of {working_dir} to {repo_id}.")
        return upload_folder(
            repo_id=repo_id, folder_path=working_dir, token=token, commit_message=commit_message, create_pr=create_pr
        )

    def push_to_hub(
        self,
        repo_id: str,
        commit_message: Optional[str] = None,
        private: Optional[bool] = None,
        token: Optional[str] = None,
        create_pr: bool = False,
        safe_serialization: bool = True,
        variant: Optional[str] = None,
    ) -> str:
        """
        Upload model, scheduler, or pipeline files to the ๐Ÿค— Hugging Face Hub.

        Parameters:
            repo_id (`str`):
                The name of the repository you want to push your model, scheduler, or pipeline files to. It should
                contain your organization name when pushing to an organization. `repo_id` can also be a path to a local
                directory.
            commit_message (`str`, *optional*):
                Message to commit while pushing. Default to `"Upload {object}"`.
            private (`bool`, *optional*):
                Whether to make the repo private. If `None` (default), the repo will be public unless the
                organization's default is private. This value is ignored if the repo already exists.
            token (`str`, *optional*):
                The token to use as HTTP bearer authorization for remote files. The token generated when running
                `huggingface-cli login` (stored in `~/.huggingface`).
            create_pr (`bool`, *optional*, defaults to `False`):
                Whether or not to create a PR with the uploaded files or directly commit.
            safe_serialization (`bool`, *optional*, defaults to `True`):
                Whether or not to convert the model weights to the `safetensors` format.
            variant (`str`, *optional*):
                If specified, weights are saved in the format `pytorch_model.<variant>.bin`.

        Examples:

        ```python
        from mindone.diffusers import UNet2DConditionModel

        unet = UNet2DConditionModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="unet")

        # Push the `unet` to your namespace with the name "my-finetuned-unet".
        unet.push_to_hub("my-finetuned-unet")

        # Push the `unet` to an organization with the name "my-finetuned-unet".
        unet.push_to_hub("your-org/my-finetuned-unet")
        ```
        """
        repo_id = create_repo(repo_id, private=private, token=token, exist_ok=True).repo_id

        # Create a new empty model card and eventually tag it
        model_card = load_or_create_model_card(repo_id, token=token)
        model_card = populate_model_card(model_card)

        # Save all files.
        save_kwargs = {"safe_serialization": safe_serialization}
        if "Scheduler" not in self.__class__.__name__:
            save_kwargs.update({"variant": variant})

        with tempfile.TemporaryDirectory() as tmpdir:
            self.save_pretrained(tmpdir, **save_kwargs)

            # Update model card if needed:
            model_card.save(os.path.join(tmpdir, "README.md"))

            return self._upload_folder(
                tmpdir,
                repo_id,
                token=token,
                commit_message=commit_message,
                create_pr=create_pr,
            )

mindone.diffusers.utils.PushToHubMixin.push_to_hub(repo_id, commit_message=None, private=None, token=None, create_pr=False, safe_serialization=True, variant=None)

Upload model, scheduler, or pipeline files to the ๐Ÿค— Hugging Face Hub.

PARAMETER DESCRIPTION
repo_id

The name of the repository you want to push your model, scheduler, or pipeline files to. It should contain your organization name when pushing to an organization. repo_id can also be a path to a local directory.

TYPE: `str`

commit_message

Message to commit while pushing. Default to "Upload {object}".

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

private

Whether to make the repo private. If None (default), the repo will be public unless the organization's default is private. This value is ignored if the repo already exists.

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

token

The token to use as HTTP bearer authorization for remote files. The token generated when running huggingface-cli login (stored in ~/.huggingface).

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

create_pr

Whether or not to create a PR with the uploaded files or directly commit.

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

safe_serialization

Whether or not to convert the model weights to the safetensors format.

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

variant

If specified, weights are saved in the format pytorch_model.<variant>.bin.

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

from mindone.diffusers import UNet2DConditionModel

unet = UNet2DConditionModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="unet")

# Push the `unet` to your namespace with the name "my-finetuned-unet".
unet.push_to_hub("my-finetuned-unet")

# Push the `unet` to an organization with the name "my-finetuned-unet".
unet.push_to_hub("your-org/my-finetuned-unet")
Source code in mindone/diffusers/utils/hub_utils.py
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
def push_to_hub(
    self,
    repo_id: str,
    commit_message: Optional[str] = None,
    private: Optional[bool] = None,
    token: Optional[str] = None,
    create_pr: bool = False,
    safe_serialization: bool = True,
    variant: Optional[str] = None,
) -> str:
    """
    Upload model, scheduler, or pipeline files to the ๐Ÿค— Hugging Face Hub.

    Parameters:
        repo_id (`str`):
            The name of the repository you want to push your model, scheduler, or pipeline files to. It should
            contain your organization name when pushing to an organization. `repo_id` can also be a path to a local
            directory.
        commit_message (`str`, *optional*):
            Message to commit while pushing. Default to `"Upload {object}"`.
        private (`bool`, *optional*):
            Whether to make the repo private. If `None` (default), the repo will be public unless the
            organization's default is private. This value is ignored if the repo already exists.
        token (`str`, *optional*):
            The token to use as HTTP bearer authorization for remote files. The token generated when running
            `huggingface-cli login` (stored in `~/.huggingface`).
        create_pr (`bool`, *optional*, defaults to `False`):
            Whether or not to create a PR with the uploaded files or directly commit.
        safe_serialization (`bool`, *optional*, defaults to `True`):
            Whether or not to convert the model weights to the `safetensors` format.
        variant (`str`, *optional*):
            If specified, weights are saved in the format `pytorch_model.<variant>.bin`.

    Examples:

    ```python
    from mindone.diffusers import UNet2DConditionModel

    unet = UNet2DConditionModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="unet")

    # Push the `unet` to your namespace with the name "my-finetuned-unet".
    unet.push_to_hub("my-finetuned-unet")

    # Push the `unet` to an organization with the name "my-finetuned-unet".
    unet.push_to_hub("your-org/my-finetuned-unet")
    ```
    """
    repo_id = create_repo(repo_id, private=private, token=token, exist_ok=True).repo_id

    # Create a new empty model card and eventually tag it
    model_card = load_or_create_model_card(repo_id, token=token)
    model_card = populate_model_card(model_card)

    # Save all files.
    save_kwargs = {"safe_serialization": safe_serialization}
    if "Scheduler" not in self.__class__.__name__:
        save_kwargs.update({"variant": variant})

    with tempfile.TemporaryDirectory() as tmpdir:
        self.save_pretrained(tmpdir, **save_kwargs)

        # Update model card if needed:
        model_card.save(os.path.join(tmpdir, "README.md"))

        return self._upload_folder(
            tmpdir,
            repo_id,
            token=token,
            commit_message=commit_message,
            create_pr=create_pr,
        )