"tests/models/vscode:/vscode.git/clone" did not exist on "c624d5ba0b41542eb3daa0e2fdc985231551d776"
modeling_utils.py 232 KB
Newer Older
1
# coding=utf-8
thomwolf's avatar
thomwolf committed
2
# Copyright 2018 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
3
4
5
6
7
8
9
10
11
12
13
14
15
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
Sylvain Gugger's avatar
Sylvain Gugger committed
16
import collections
17
import copy
18
import functools
19
import gc
20
import importlib.metadata
Yih-Dar's avatar
Yih-Dar committed
21
import inspect
22
import itertools
Sylvain Gugger's avatar
Sylvain Gugger committed
23
import json
24
import os
25
import re
Sylvain Gugger's avatar
Sylvain Gugger committed
26
27
import shutil
import tempfile
28
import warnings
29
from contextlib import contextmanager
30
from dataclasses import dataclass
31
from functools import partial, wraps
32
from threading import Thread
33
from typing import Any, Callable, Dict, List, Optional, Set, Tuple, Union
34
from zipfile import is_zipfile
35
36

import torch
37
from packaging import version
Sylvain Gugger's avatar
Sylvain Gugger committed
38
from torch import Tensor, nn
39
from torch.nn import CrossEntropyLoss, Identity
40
from torch.utils.checkpoint import checkpoint
41

42
from .activations import get_activation
43
from .configuration_utils import PretrainedConfig
44
from .dynamic_module_utils import custom_object_save
45
from .generation import GenerationConfig, GenerationMixin
46
from .integrations import PeftAdapterMixin, deepspeed_config, is_deepspeed_zero3_enabled
47
48
49
50
from .pytorch_utils import (  # noqa: F401
    Conv1D,
    apply_chunking_to_forward,
    find_pruneable_heads_and_indices,
Thomas Wang's avatar
Thomas Wang committed
51
    id_tensor_storage,
52
    is_torch_greater_or_equal_than_1_13,
53
54
55
56
    prune_conv1d_layer,
    prune_layer,
    prune_linear_layer,
)
57
from .quantizers import AutoHfQuantizer, HfQuantizer
58
from .quantizers.quantizers_utils import get_module_from_name
59
from .safetensors_conversion import auto_conversion
60
from .utils import (
61
62
    ADAPTER_SAFE_WEIGHTS_NAME,
    ADAPTER_WEIGHTS_NAME,
63
    CONFIG_NAME,
Aymeric Augustin's avatar
Aymeric Augustin committed
64
    DUMMY_INPUTS,
65
    FLAX_WEIGHTS_NAME,
66
67
    SAFE_WEIGHTS_INDEX_NAME,
    SAFE_WEIGHTS_NAME,
68
69
    TF2_WEIGHTS_NAME,
    TF_WEIGHTS_NAME,
Sylvain Gugger's avatar
Sylvain Gugger committed
70
    WEIGHTS_INDEX_NAME,
71
    WEIGHTS_NAME,
72
    ContextManagers,
73
    ModelOutput,
Sylvain Gugger's avatar
Sylvain Gugger committed
74
    PushToHubMixin,
75
    cached_file,
76
    copy_func,
77
    download_url,
78
    extract_commit_hash,
79
    has_file,
80
    is_accelerate_available,
81
    is_bitsandbytes_available,
82
    is_flash_attn_2_available,
83
    is_offline_mode,
84
    is_optimum_available,
85
    is_peft_available,
86
    is_remote_url,
87
    is_safetensors_available,
88
    is_torch_sdpa_available,
89
    is_torch_xla_available,
90
    logging,
Sylvain Gugger's avatar
Sylvain Gugger committed
91
    replace_return_docstrings,
92
    strtobool,
93
)
94
from .utils.hub import convert_file_size_to_int, create_and_tag_model_card, get_checkpoint_shard_files
95
96
97
98
99
100
from .utils.import_utils import (
    ENV_VARS_TRUE_VALUES,
    is_sagemaker_mp_enabled,
    is_torch_fx_proxy,
    is_torchdynamo_compiling,
)
101
from .utils.quantization_config import BitsAndBytesConfig, QuantizationMethod
102

Aymeric Augustin's avatar
Aymeric Augustin committed
103

104
105
106
XLA_USE_BF16 = os.environ.get("XLA_USE_BF16", "0").upper()
XLA_DOWNCAST_BF16 = os.environ.get("XLA_DOWNCAST_BF16", "0").upper()

107
108
if is_accelerate_available():
    from accelerate import dispatch_model, infer_auto_device_map, init_empty_weights
109
    from accelerate.hooks import add_hook_to_module
110
    from accelerate.utils import (
111
        check_tied_parameters_on_same_device,
112
        find_tied_parameters,
113
        get_balanced_memory,
Marc Sun's avatar
Marc Sun committed
114
        get_max_memory,
115
116
117
118
119
120
        load_offloaded_weights,
        offload_weight,
        save_offload_index,
        set_module_tensor_to_device,
    )

121
122
123
124
if is_safetensors_available():
    from safetensors import safe_open
    from safetensors.torch import load_file as safe_load_file
    from safetensors.torch import save_file as safe_save_file
125

Lysandre Debut's avatar
Lysandre Debut committed
126
logger = logging.get_logger(__name__)
127

128
129
130
131

_init_weights = True


132
def is_fsdp_enabled():
133
134
135
136
    return (
        torch.distributed.is_available()
        and torch.distributed.is_initialized()
        and strtobool(os.environ.get("ACCELERATE_USE_FSDP", "False")) == 1
137
        and strtobool(os.environ.get("FSDP_CPU_RAM_EFFICIENT_LOADING", "False")) == 1
138
    )
139
140


141
142
143
144
145
146
def is_local_dist_rank_0():
    return (
        torch.distributed.is_available()
        and torch.distributed.is_initialized()
        and int(os.environ.get("LOCAL_RANK", -1)) == 0
    )
147
148


149
150
151
152
153
154
155
156
if is_sagemaker_mp_enabled():
    import smdistributed.modelparallel.torch as smp
    from smdistributed.modelparallel import __version__ as SMP_VERSION

    IS_SAGEMAKER_MP_POST_1_10 = version.parse(SMP_VERSION) >= version.parse("1.10")
else:
    IS_SAGEMAKER_MP_POST_1_10 = False

157
158
159
if is_peft_available():
    from .utils import find_adapter_config_file

160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
TORCH_INIT_FUNCTIONS = {
    "uniform_": nn.init.uniform_,
    "normal_": nn.init.normal_,
    "trunc_normal_": nn.init.trunc_normal_,
    "constant_": nn.init.constant_,
    "xavier_uniform_": nn.init.xavier_uniform_,
    "xavier_normal_": nn.init.xavier_normal_,
    "kaiming_uniform_": nn.init.kaiming_uniform_,
    "kaiming_normal_": nn.init.kaiming_normal_,
    "uniform": nn.init.uniform,
    "normal": nn.init.normal,
    "xavier_uniform": nn.init.xavier_uniform,
    "xavier_normal": nn.init.xavier_normal,
    "kaiming_uniform": nn.init.kaiming_uniform,
    "kaiming_normal": nn.init.kaiming_normal,
}

177

178
179
180
181
182
183
184
185
@contextmanager
def no_init_weights(_enable=True):
    """
    Context manager to globally disable weight initialization to speed up loading large models.

    TODO(Patrick): Delete safety argument `_enable=True` at next major version. .
    """
    global _init_weights
186
    old_init_weights = _init_weights
187

188
189
    if _enable:
        _init_weights = False
190
191
192
193
194
195
196

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

        # # Save the original initialization functions
        for name, init_func in TORCH_INIT_FUNCTIONS.items():
            setattr(torch.nn.init, name, _skip_init)
197
198
199
    try:
        yield
    finally:
200
        _init_weights = old_init_weights
201
202
203
204
        if _enable:
            # # Restore the original initialization functions
            for name, init_func in TORCH_INIT_FUNCTIONS.items():
                setattr(torch.nn.init, name, init_func)
205
206


Lysandre Debut's avatar
Lysandre Debut committed
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
def get_parameter_device(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
    try:
        return next(parameter.parameters()).device
    except StopIteration:
        # For nn.DataParallel compatibility in PyTorch 1.5

        def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
            tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
            return tuples

        gen = parameter._named_members(get_members_fn=find_tensor_attributes)
        first_tuple = next(gen)
        return first_tuple[1].device


222
223
224
225
def get_first_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
    """
    Returns the first parameter dtype (can be non-floating) or asserts if none were found.
    """
Lysandre Debut's avatar
Lysandre Debut committed
226
227
228
    try:
        return next(parameter.parameters()).dtype
    except StopIteration:
Sylvain Gugger's avatar
Sylvain Gugger committed
229
        # For nn.DataParallel compatibility in PyTorch > 1.5
Lysandre Debut's avatar
Lysandre Debut committed
230
231
232
233
234
235
236
237
238
239

        def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
            tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
            return tuples

        gen = parameter._named_members(get_members_fn=find_tensor_attributes)
        first_tuple = next(gen)
        return first_tuple[1].dtype


240
241
def get_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
242
    Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.
243
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
244
245
246
247
    last_dtype = None
    for t in parameter.parameters():
        last_dtype = t.dtype
        if t.is_floating_point():
248
249
250
            # Adding fix for https://github.com/pytorch/xla/issues/4152
            # Fixes issue where the model code passes a value that is out of range for XLA_USE_BF16=1
            # and XLA_DOWNCAST_BF16=1 so the conversion would cast it to -inf
251
252
            # NOTE: `is_torch_xla_available()` is checked last as it induces a graph break in torch dynamo
            if XLA_USE_BF16 in ENV_VARS_TRUE_VALUES and is_torch_xla_available():
253
                return torch.bfloat16
254
            if XLA_DOWNCAST_BF16 in ENV_VARS_TRUE_VALUES and is_torch_xla_available():
255
                if t.dtype == torch.float:
256
                    return torch.bfloat16
257
258
                if t.dtype == torch.double:
                    return torch.float32
Sylvain Gugger's avatar
Sylvain Gugger committed
259
            return t.dtype
260

Sylvain Gugger's avatar
Sylvain Gugger committed
261
262
263
    if last_dtype is not None:
        # if no floating dtype was found return whatever the first dtype is
        return last_dtype
264

265
266
267
268
269
270
271
272
273
274
275
276
277
    # For nn.DataParallel compatibility in PyTorch > 1.5
    def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
        tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
        return tuples

    gen = parameter._named_members(get_members_fn=find_tensor_attributes)
    last_tuple = None
    for tuple in gen:
        last_tuple = tuple
        if tuple[1].is_floating_point():
            return tuple[1].dtype

    if last_tuple is not None:
278
279
        # fallback to the last dtype
        return last_tuple[1].dtype
280

281
282
283
284
285
286
287
    # fallback to buffer dtype
    for t in parameter.buffers():
        last_dtype = t.dtype
        if t.is_floating_point():
            return t.dtype
    return last_dtype

288
289
290
291
292
293
294
295
296
297
298
299
300
301

def get_state_dict_float_dtype(state_dict):
    """
    Returns the first found floating dtype in `state_dict` or asserts if none were found.
    """
    for t in state_dict.values():
        if t.is_floating_point():
            return t.dtype

    raise ValueError("couldn't find any floating point dtypes in state_dict")


def get_state_dict_dtype(state_dict):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
302
    Returns the first found floating dtype in `state_dict` if there is one, otherwise returns the first dtype.
303
304
305
306
307
308
309
    """
    for t in state_dict.values():
        if t.is_floating_point():
            return t.dtype

    # if no floating dtype was found return whatever the first dtype is
    else:
Sylvain Gugger's avatar
Sylvain Gugger committed
310
        return next(state_dict.values()).dtype
311
312


Sylvain Gugger's avatar
Sylvain Gugger committed
313
314
315
316
317
318
319
320
321
322
323
324
325
def dtype_byte_size(dtype):
    """
    Returns the size (in bytes) occupied by one parameter of type `dtype`.

    Example:

    ```py
    >>> dtype_byte_size(torch.float32)
    4
    ```
    """
    if dtype == torch.bool:
        return 1 / 8
326
    bit_search = re.search(r"[^\d](\d+)$", str(dtype))
Sylvain Gugger's avatar
Sylvain Gugger committed
327
328
329
330
331
332
    if bit_search is None:
        raise ValueError(f"`dtype` is not a valid dtype: {dtype}.")
    bit_size = int(bit_search.groups()[0])
    return bit_size // 8


333
334
335
def shard_checkpoint(
    state_dict: Dict[str, torch.Tensor], max_shard_size: Union[int, str] = "10GB", weights_name: str = WEIGHTS_NAME
):
Sylvain Gugger's avatar
Sylvain Gugger committed
336
337
338
339
340
341
342
343
344
345
346
    """
    Splits a model state dictionary in sub-checkpoints so that the final size of each sub-checkpoint does not exceed a
    given size.

    The sub-checkpoints are determined by iterating through the `state_dict` in the order of its keys, so there is no
    optimization made to make each sub-checkpoint as close as possible to the maximum size passed. For example, if the
    limit is 10GB and we have weights of sizes [6GB, 6GB, 2GB, 6GB, 2GB, 2GB] they will get sharded as [6GB], [6+2GB],
    [6+2+2GB] and not [6+2+2GB], [6+2GB], [6GB].

    <Tip warning={true}>

Alan Ji's avatar
Alan Ji committed
347
    If one of the model's weight is bigger than `max_shard_size`, it will end up in its own sub-checkpoint which will
Sylvain Gugger's avatar
Sylvain Gugger committed
348
349
350
351
352
353
354
355
356
    have a size greater than `max_shard_size`.

    </Tip>

    Args:
        state_dict (`Dict[str, torch.Tensor]`): The state dictionary of a model to save.
        max_shard_size (`int` or `str`, *optional*, defaults to `"10GB"`):
            The maximum size of each sub-checkpoint. If expressed as a string, needs to be digits followed by a unit
            (like `"5MB"`).
357
358
        weights_name (`str`, *optional*, defaults to `"pytorch_model.bin"`):
            The name of the model save file.
Sylvain Gugger's avatar
Sylvain Gugger committed
359
360
361
    """
    max_shard_size = convert_file_size_to_int(max_shard_size)

Thomas Wang's avatar
Thomas Wang committed
362
363
    sharded_state_dicts = [{}]
    last_block_size = 0
Sylvain Gugger's avatar
Sylvain Gugger committed
364
    total_size = 0
Thomas Wang's avatar
Thomas Wang committed
365
    storage_id_to_block = {}
Sylvain Gugger's avatar
Sylvain Gugger committed
366
367

    for key, weight in state_dict.items():
368
369
370
371
372
373
        # when bnb serialization is used the weights in the state dict can be strings
        # check: https://github.com/huggingface/transformers/pull/24416 for more details
        if isinstance(weight, str):
            continue
        else:
            storage_id = id_tensor_storage(weight)
Thomas Wang's avatar
Thomas Wang committed
374
375
376
377
378
379
380

        # If a `weight` shares the same underlying storage as another tensor, we put `weight` in the same `block`
        if storage_id in storage_id_to_block:
            block_id = storage_id_to_block[storage_id]
            sharded_state_dicts[block_id][key] = weight
            continue

Sylvain Gugger's avatar
Sylvain Gugger committed
381
382
        weight_size = weight.numel() * dtype_byte_size(weight.dtype)

Sylvain Gugger's avatar
Sylvain Gugger committed
383
384
385
        # If this weight is going to tip up over the maximal size, we split, but only if we have put at least one
        # weight in the current shard.
        if last_block_size + weight_size > max_shard_size and len(sharded_state_dicts[-1]) > 0:
Thomas Wang's avatar
Thomas Wang committed
386
387
            sharded_state_dicts.append({})
            last_block_size = 0
Sylvain Gugger's avatar
Sylvain Gugger committed
388

Thomas Wang's avatar
Thomas Wang committed
389
390
        sharded_state_dicts[-1][key] = weight
        last_block_size += weight_size
Sylvain Gugger's avatar
Sylvain Gugger committed
391
        total_size += weight_size
Thomas Wang's avatar
Thomas Wang committed
392
        storage_id_to_block[storage_id] = len(sharded_state_dicts) - 1
Sylvain Gugger's avatar
Sylvain Gugger committed
393
394
395

    # If we only have one shard, we return it
    if len(sharded_state_dicts) == 1:
396
        return {weights_name: sharded_state_dicts[0]}, None
Sylvain Gugger's avatar
Sylvain Gugger committed
397
398
399
400
401

    # Otherwise, let's build the index
    weight_map = {}
    shards = {}
    for idx, shard in enumerate(sharded_state_dicts):
402
403
404
405
        shard_file = weights_name.replace(".bin", f"-{idx+1:05d}-of-{len(sharded_state_dicts):05d}.bin")
        shard_file = shard_file.replace(
            ".safetensors", f"-{idx + 1:05d}-of-{len(sharded_state_dicts):05d}.safetensors"
        )
Sylvain Gugger's avatar
Sylvain Gugger committed
406
407
408
409
410
411
412
413
414
415
        shards[shard_file] = shard
        for key in shard.keys():
            weight_map[key] = shard_file

    # Add the metadata
    metadata = {"total_size": total_size}
    index = {"metadata": metadata, "weight_map": weight_map}
    return shards, index


416
def load_sharded_checkpoint(model, folder, strict=True, prefer_safe=True):
417
418
419
420
421
422
423
424
425
426
427
428
429
    """
    This is the same as
    [`torch.nn.Module.load_state_dict`](https://pytorch.org/docs/stable/generated/torch.nn.Module.html?highlight=load_state_dict#torch.nn.Module.load_state_dict)
    but for a sharded checkpoint.

    This load is performed efficiently: each checkpoint shard is loaded one by one in RAM and deleted after being
    loaded in the model.

    Args:
        model (`torch.nn.Module`): The model in which to load the checkpoint.
        folder (`str` or `os.PathLike`): A path to a folder containing the sharded checkpoint.
        strict (`bool`, *optional`, defaults to `True`):
            Whether to strictly enforce that the keys in the model state dict match the keys in the sharded checkpoint.
430
431
432
        prefer_safe (`bool`, *optional*, defaults to `False`)
            If both safetensors and PyTorch save files are present in checkpoint and `prefer_safe` is True, the
            safetensors files will be loaded. Otherwise, PyTorch files are always loaded when possible.
433
434
435
436
437
438
439
440

    Returns:
        `NamedTuple`: A named tuple with `missing_keys` and `unexpected_keys` fields
            - `missing_keys` is a list of str containing the missing keys
            - `unexpected_keys` is a list of str containing the unexpected keys
    """
    # Load the index
    index_file = os.path.join(folder, WEIGHTS_INDEX_NAME)
441
    safe_index_file = os.path.join(folder, SAFE_WEIGHTS_INDEX_NAME)
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
    index_present = os.path.isfile(index_file)
    safe_index_present = os.path.isfile(safe_index_file)

    if not index_present and not (safe_index_present and is_safetensors_available()):
        filenames = (
            (WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_INDEX_NAME) if is_safetensors_available() else (WEIGHTS_INDEX_NAME,)
        )
        raise ValueError(f"Can't find a checkpoint index ({' or '.join(filenames)}) in {folder}.")

    load_safe = False
    if safe_index_present:
        if prefer_safe:
            if is_safetensors_available():
                load_safe = True  # load safe due to preference
            else:
                logger.warning(
                    f"Cannot load sharded checkpoint at {folder} safely since safetensors is not installed!"
                )
        elif not index_present:
            load_safe = True  # load safe since we have no other choice

    load_index = safe_index_file if load_safe else index_file

    with open(load_index, "r", encoding="utf-8") as f:
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
        index = json.load(f)

    shard_files = list(set(index["weight_map"].values()))

    # If strict=True, error before loading any of the state dicts.
    loaded_keys = index["weight_map"].keys()
    model_keys = model.state_dict().keys()
    missing_keys = [key for key in model_keys if key not in loaded_keys]
    unexpected_keys = [key for key in loaded_keys if key not in model_keys]
    if strict and (len(missing_keys) > 0 or len(unexpected_keys) > 0):
        error_message = f"Error(s) in loading state_dict for {model.__class__.__name__}"
        if len(missing_keys) > 0:
            str_missing_keys = ",".join([f'"{k}"' for k in missing_keys])
            error_message += f"\nMissing key(s): {str_missing_keys}."
        if len(unexpected_keys) > 0:
            str_unexpected_keys = ",".join([f'"{k}"' for k in unexpected_keys])
            error_message += f"\nMissing key(s): {str_unexpected_keys}."
        raise RuntimeError(error_message)

Yih-Dar's avatar
Yih-Dar committed
486
487
    weights_only_kwarg = {"weights_only": True} if is_torch_greater_or_equal_than_1_13 else {}
    loader = safe_load_file if load_safe else partial(torch.load, map_location="cpu", **weights_only_kwarg)
488

489
    for shard_file in shard_files:
490
        state_dict = loader(os.path.join(folder, shard_file))
491
492
        model.load_state_dict(state_dict, strict=False)

493
        # Make sure memory is freed before we load the next state dict.
494
495
496
497
498
499
500
        del state_dict
        gc.collect()

    # Return the same thing as PyTorch load_state_dict function.
    return torch.nn.modules.module._IncompatibleKeys(missing_keys, unexpected_keys)


501
def load_state_dict(checkpoint_file: Union[str, os.PathLike], is_quantized: bool = False):
Sylvain Gugger's avatar
Sylvain Gugger committed
502
503
504
    """
    Reads a PyTorch checkpoint file, returning properly formatted errors if they arise.
    """
505
506
507
508
    if checkpoint_file.endswith(".safetensors") and is_safetensors_available():
        # Check format of the archive
        with safe_open(checkpoint_file, framework="pt") as f:
            metadata = f.metadata()
509
        if metadata.get("format") not in ["pt", "tf", "flax", "mlx"]:
510
511
512
513
514
            raise OSError(
                f"The safetensors archive passed at {checkpoint_file} does not contain the valid metadata. Make sure "
                "you save your model with the `save_pretrained` method."
            )
        return safe_load_file(checkpoint_file)
Sylvain Gugger's avatar
Sylvain Gugger committed
515
    try:
516
        if (
517
518
519
            (is_deepspeed_zero3_enabled() and torch.distributed.is_initialized() and torch.distributed.get_rank() > 0)
            or (is_fsdp_enabled() and not is_local_dist_rank_0())
        ) and not is_quantized:
520
521
522
            map_location = "meta"
        else:
            map_location = "cpu"
523
524
525
526
527
528
529
530
531
        extra_args = {}
        # mmap can only be used with files serialized with zipfile-based format.
        if (
            isinstance(checkpoint_file, str)
            and map_location != "meta"
            and version.parse(torch.__version__) >= version.parse("2.1.0")
            and is_zipfile(checkpoint_file)
        ):
            extra_args = {"mmap": True}
Yih-Dar's avatar
Yih-Dar committed
532
        weights_only_kwarg = {"weights_only": True} if is_torch_greater_or_equal_than_1_13 else {}
533
534
535
        return torch.load(
            checkpoint_file,
            map_location=map_location,
Yih-Dar's avatar
Yih-Dar committed
536
            **weights_only_kwarg,
537
538
            **extra_args,
        )
Sylvain Gugger's avatar
Sylvain Gugger committed
539
540
541
    except Exception as e:
        try:
            with open(checkpoint_file) as f:
542
                if f.read(7) == "version":
Sylvain Gugger's avatar
Sylvain Gugger committed
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
                    raise OSError(
                        "You seem to have cloned a repository without having git-lfs installed. Please install "
                        "git-lfs and run `git lfs install` followed by `git lfs pull` in the folder "
                        "you cloned."
                    )
                else:
                    raise ValueError(
                        f"Unable to locate the file {checkpoint_file} which is necessary to load this pretrained "
                        "model. Make sure you have saved the model properly."
                    ) from e
        except (UnicodeDecodeError, ValueError):
            raise OSError(
                f"Unable to load weights from pytorch checkpoint file for '{checkpoint_file}' "
                f"at '{checkpoint_file}'. "
                "If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True."
            )


561
562
563
564
565
def set_initialized_submodules(model, state_dict_keys):
    """
    Sets the `_is_hf_initialized` flag in all submodules of a given model when all its weights are in the loaded state
    dict.
    """
566
    not_initialized_submodules = {}
567
    for module_name, module in model.named_modules():
568
569
        loaded_keys = {k.replace(f"{module_name}.", "") for k in state_dict_keys if k.startswith(f"{module_name}.")}
        if loaded_keys.issuperset(module.state_dict()):
570
            module._is_hf_initialized = True
571
572
573
        else:
            not_initialized_submodules[module_name] = module
    return not_initialized_submodules
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
def _end_ptr(tensor: torch.Tensor) -> int:
    # extract the end of the pointer if the tensor is a slice of a bigger tensor
    if tensor.nelement():
        stop = tensor.view(-1)[-1].data_ptr() + tensor.element_size()
    else:
        stop = tensor.data_ptr()
    return stop


def _get_tied_weight_keys(module: nn.Module, prefix=""):
    tied_weight_keys = []
    if getattr(module, "_tied_weights_keys", None) is not None:
        names = [f"{prefix}.{k}" if prefix else k for k in module._tied_weights_keys]
        tied_weight_keys.extend(names)
    if getattr(module, "_dynamic_tied_weights_keys", None) is not None:
        names = [f"{prefix}.{k}" if prefix else k for k in module._dynamic_tied_weights_keys]
        tied_weight_keys.extend(names)
    for name, submodule in module.named_children():
        local_prefix = f"{prefix}.{name}" if prefix else name
        tied_weight_keys.extend(_get_tied_weight_keys(submodule, prefix=local_prefix))
    return tied_weight_keys


def _find_disjoint(tensors: List[Set[str]], state_dict: Dict[str, torch.Tensor]) -> Tuple[List[Set[str]], List[str]]:
    filtered_tensors = []
    for shared in tensors:
        if len(shared) < 2:
            filtered_tensors.append(shared)
            continue

        areas = []
        for name in shared:
            tensor = state_dict[name]
            areas.append((tensor.data_ptr(), _end_ptr(tensor), name))
        areas.sort()

        _, last_stop, last_name = areas[0]
        filtered_tensors.append({last_name})
        for start, stop, name in areas[1:]:
            if start >= last_stop:
                filtered_tensors.append({name})
            else:
                filtered_tensors[-1].add(name)
            last_stop = stop
    disjoint_tensors = []
    shared_tensors = []
    for tensors in filtered_tensors:
        if len(tensors) == 1:
            disjoint_tensors.append(tensors.pop())
        else:
            shared_tensors.append(tensors)
    return shared_tensors, disjoint_tensors


def _find_identical(tensors: List[Set[str]], state_dict: Dict[str, torch.Tensor]) -> Tuple[List[Set[str]], Set[str]]:
    shared_tensors = []
    identical = []
    for shared in tensors:
        if len(shared) < 2:
            continue

        areas = collections.defaultdict(set)
        for name in shared:
            tensor = state_dict[name]
            area = (tensor.device, tensor.data_ptr(), _end_ptr(tensor))
            areas[area].add(name)
        if len(areas) == 1:
            identical.append(shared)
        else:
            shared_tensors.append(shared)
    return shared_tensors, identical


Sylvain Gugger's avatar
Sylvain Gugger committed
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
def _load_state_dict_into_model(model_to_load, state_dict, start_prefix):
    # Convert old format to new format if needed from a PyTorch state_dict
    old_keys = []
    new_keys = []
    for key in state_dict.keys():
        new_key = None
        if "gamma" in key:
            new_key = key.replace("gamma", "weight")
        if "beta" in key:
            new_key = key.replace("beta", "bias")
        if new_key:
            old_keys.append(key)
            new_keys.append(new_key)
    for old_key, new_key in zip(old_keys, new_keys):
        state_dict[new_key] = state_dict.pop(old_key)

    # copy state_dict so _load_from_state_dict can modify it
    metadata = getattr(state_dict, "_metadata", None)
    state_dict = state_dict.copy()
    if metadata is not None:
        state_dict._metadata = metadata

    error_msgs = []

    # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants
    # so we need to apply the function recursively.
675
    def load(module: nn.Module, state_dict, prefix=""):
Sylvain Gugger's avatar
Sylvain Gugger committed
676
        local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
677
        args = (state_dict, prefix, local_metadata, True, [], [], error_msgs)
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
        # Parameters of module and children will start with prefix. We can exit early if there are none in this
        # state_dict
        if len([key for key in state_dict if key.startswith(prefix)]) > 0:
            if is_deepspeed_zero3_enabled():
                import deepspeed

                # In sharded models, each shard has only part of the full state_dict, so only gather
                # parameters that are in the current state_dict.
                named_parameters = dict(module.named_parameters(prefix=prefix[:-1], recurse=False))
                params_to_gather = [named_parameters[k] for k in state_dict.keys() if k in named_parameters]
                if len(params_to_gather) > 0:
                    # because zero3 puts placeholders in model params, this context
                    # manager gathers (unpartitions) the params of the current layer, then loads from
                    # the state dict and then re-partitions them again
                    with deepspeed.zero.GatheredParameters(params_to_gather, modifier_rank=0):
                        if torch.distributed.get_rank() == 0:
                            module._load_from_state_dict(*args)
            else:
                module._load_from_state_dict(*args)
Sylvain Gugger's avatar
Sylvain Gugger committed
697
698
699

        for name, child in module._modules.items():
            if child is not None:
700
                load(child, state_dict, prefix + name + ".")
Sylvain Gugger's avatar
Sylvain Gugger committed
701

702
703
704
705
    load(model_to_load, state_dict, prefix=start_prefix)
    # Delete `state_dict` so it could be collected by GC earlier. Note that `state_dict` is a copy of the argument, so
    # it's safe to delete it.
    del state_dict
Sylvain Gugger's avatar
Sylvain Gugger committed
706
707
708
709

    return error_msgs


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
def find_submodule_and_param_name(model, long_key, start_prefix):
    """
    A helper util to find the last sub-module and the param/buffer name. If `start_prefix` is supplied it'll be removed
    from the start of the key
    """

    if len(start_prefix) > 0 and long_key.startswith(start_prefix):
        long_key = ".".join(long_key.split(".")[1:])

    split_key = long_key.split(".")
    submodule = model
    while len(split_key) > 1:
        if hasattr(submodule, split_key[0]):
            submodule = getattr(submodule, split_key[0])
            del split_key[0]
        else:
            submodule = None
            break
    if submodule == model:
        submodule = None
    return submodule, split_key[0]


def _move_model_to_meta(model, loaded_state_dict_keys, start_prefix):
    """
    Moves `loaded_state_dict_keys` in model to meta device which frees up the memory taken by those params.

    `start_prefix` is used for models which insert their name into model keys, e.g. `bert` in
    `bert.pooler.dense.weight`

    """

    # dematerialize param storage for keys that are going to be replaced by state_dict, by
    # putting those on the meta device
    for k in loaded_state_dict_keys:
        submodule, param_name = find_submodule_and_param_name(model, k, start_prefix)
        if submodule is not None:
            # selectively switch to the meta device only those params/buffers that will
            # be next replaced from state_dict. This a complex way to do p.to_("meta")
            # since we have no in-place to_ for tensors.
            new_val = getattr(submodule, param_name)
            if isinstance(new_val, torch.nn.Parameter):
                # isinstance returns False for Params on meta device, so switch after the check
                new_val = torch.nn.Parameter(new_val.to("meta"))
            else:
                new_val = new_val.to("meta")
            setattr(submodule, param_name, new_val)


759
760
761
762
763
764
765
766
767
768
769
770
def _load_state_dict_into_meta_model(
    model,
    state_dict,
    loaded_state_dict_keys,  # left for now but could be removed, see below
    start_prefix,
    expected_keys,
    device_map=None,
    offload_folder=None,
    offload_index=None,
    state_dict_folder=None,
    state_dict_index=None,
    dtype=None,
771
    hf_quantizer=None,
Sylvain Gugger's avatar
Sylvain Gugger committed
772
    is_safetensors=False,
773
    keep_in_fp32_modules=None,
774
    unexpected_keys=None,  # passing `unexpected` for cleanup from quantization items
775
):
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
    """
    This is somewhat similar to `_load_state_dict_into_model`, but deals with a model that has some or all of its
    params on a `meta` device. It replaces the model params with the data from the `state_dict`, while moving the
    params back to the normal device, but only for `loaded_state_dict_keys`.

    `start_prefix` is used for models which insert their name into model keys, e.g. `bert` in
    `bert.pooler.dense.weight`

    """

    # XXX: remaining features to implement to be fully compatible with _load_state_dict_into_model
    # - deepspeed zero 3 support
    # - need to copy metadata if any - see _load_state_dict_into_model
    # - handling error_msgs - mimicking the error handling in module._load_from_state_dict()
    # - Is there a situation where some keys aren't in `loaded_state_dict_keys` and in which case
    #   they won't get loaded.

    error_msgs = []

795
796
    old_keys = []
    new_keys = []
797
    is_quantized = hf_quantizer is not None
798
799
800
801
802
803
804
805
806
807
808
    for key in state_dict.keys():
        new_key = None
        if "gamma" in key:
            new_key = key.replace("gamma", "weight")
        if "beta" in key:
            new_key = key.replace("beta", "bias")
        if new_key:
            old_keys.append(key)
            new_keys.append(new_key)
    for old_key, new_key in zip(old_keys, new_keys):
        state_dict[new_key] = state_dict.pop(old_key)
809

810
811
812
813
814
815
816
817
818
    for param_name, param in state_dict.items():
        # First part of the test is always true as load_state_dict_keys always contains state_dict keys.
        if param_name not in loaded_state_dict_keys or param_name not in expected_keys:
            continue

        if param_name.startswith(start_prefix):
            param_name = param_name[len(start_prefix) :]

        module_name = param_name
819
        set_module_kwargs = {}
820

821
        # We convert floating dtypes to the `dtype` passed. We want to keep the buffers/params
822
823
        # in int/uint/bool and not cast them.
        if dtype is not None and torch.is_floating_point(param):
824
825
            if (
                keep_in_fp32_modules is not None
826
827
828
                and any(
                    module_to_keep_in_fp32 in param_name.split(".") for module_to_keep_in_fp32 in keep_in_fp32_modules
                )
829
830
831
                and dtype == torch.float16
            ):
                param = param.to(torch.float32)
832
833
834
835
836

                # For backward compatibility with older versions of `accelerate`
                # TODO: @sgugger replace this check with version check at the next `accelerate` release
                if "dtype" in list(inspect.signature(set_module_tensor_to_device).parameters):
                    set_module_kwargs["dtype"] = torch.float32
837
838
            else:
                param = param.to(dtype)
839

840
841
842
843
844
845
846
847
848
849
850
851
        # For compatibility with PyTorch load_state_dict which converts state dict dtype to existing dtype in model, and which
        # uses `param.copy_(input_param)` that preserves the contiguity of the parameter in the model.
        # Reference: https://github.com/pytorch/pytorch/blob/db79ceb110f6646523019a59bbd7b838f43d4a86/torch/nn/modules/module.py#L2040C29-L2040C29
        old_param = model
        splits = param_name.split(".")
        for split in splits:
            old_param = getattr(old_param, split)
            if old_param is None:
                break

        if old_param is not None:
            if dtype is None:
852
                param = param.to(old_param.dtype)
853

854
855
856
            if old_param.is_contiguous():
                param = param.contiguous()

857
858
        set_module_kwargs["value"] = param

859
860
861
862
863
864
865
866
867
868
869
        if device_map is None:
            param_device = "cpu"
        else:
            # find next higher level module that is defined in device_map:
            # bert.lm_head.weight -> bert.lm_head -> bert -> ''
            while len(module_name) > 0 and module_name not in device_map:
                module_name = ".".join(module_name.split(".")[:-1])
            if module_name == "" and "" not in device_map:
                # TODO: group all errors and raise at the end.
                raise ValueError(f"{param_name} doesn't have any device set.")
            param_device = device_map[module_name]
870

871
        if param_device == "disk":
Sylvain Gugger's avatar
Sylvain Gugger committed
872
873
            if not is_safetensors:
                offload_index = offload_weight(param, param_name, offload_folder, offload_index)
874
        elif param_device == "cpu" and state_dict_index is not None:
875
            state_dict_index = offload_weight(param, param_name, state_dict_folder, state_dict_index)
876
        elif (
877
            not is_quantized
878
            or (not hf_quantizer.requires_parameters_quantization)
879
880
881
882
883
            or (
                not hf_quantizer.check_quantized_param(
                    model, param, param_name, state_dict, param_device=param_device, device_map=device_map
                )
            )
884
885
        ):
            # For backward compatibility with older versions of `accelerate` and for non-quantized params
886
            set_module_tensor_to_device(model, param_name, param_device, **set_module_kwargs)
887
        else:
888
            hf_quantizer.create_quantized_param(model, param, param_name, param_device, state_dict, unexpected_keys)
889
890
891
892
893
894
895
896
            # For quantized modules with FSDP/DeepSpeed Stage 3, we need to quantize the parameter on the GPU
            # and then cast it to CPU to avoid excessive memory usage on each GPU
            # in comparison to the sharded model across GPUs.
            if is_fsdp_enabled() or is_deepspeed_zero3_enabled():
                module, tensor_name = get_module_from_name(model, param_name)
                value = getattr(module, tensor_name)
                value = type(value)(value.data.to("cpu"), **value.__dict__)
                setattr(module, tensor_name, value)
897
            # TODO: consider removing used param_parts from state_dict before return
898
899

    return error_msgs, offload_index, state_dict_index
900
901


902
903
904
905
906
907
908
909
910
def _add_variant(weights_name: str, variant: Optional[str] = None) -> str:
    if variant is not None:
        splits = weights_name.split(".")
        splits = splits[:-1] + [variant] + splits[-1:]
        weights_name = ".".join(splits)

    return weights_name


911
class ModuleUtilsMixin:
Julien Chaumond's avatar
Julien Chaumond committed
912
    """
913
    A few utilities for `torch.nn.Modules`, to be used as a mixin.
Julien Chaumond's avatar
Julien Chaumond committed
914
915
    """

916
917
918
919
    @staticmethod
    def _hook_rss_memory_pre_forward(module, *args, **kwargs):
        try:
            import psutil
Sylvain Gugger's avatar
Sylvain Gugger committed
920
        except ImportError:
921
922
923
924
925
926
927
928
929
930
931
            raise ImportError("You need to install psutil (pip install psutil) to use memory tracing.")

        process = psutil.Process(os.getpid())
        mem = process.memory_info()
        module.mem_rss_pre_forward = mem.rss
        return None

    @staticmethod
    def _hook_rss_memory_post_forward(module, *args, **kwargs):
        try:
            import psutil
Sylvain Gugger's avatar
Sylvain Gugger committed
932
        except ImportError:
933
934
935
936
937
938
939
940
941
942
            raise ImportError("You need to install psutil (pip install psutil) to use memory tracing.")

        process = psutil.Process(os.getpid())
        mem = process.memory_info()
        module.mem_rss_post_forward = mem.rss
        mem_rss_diff = module.mem_rss_post_forward - module.mem_rss_pre_forward
        module.mem_rss_diff = mem_rss_diff + (module.mem_rss_diff if hasattr(module, "mem_rss_diff") else 0)
        return None

    def add_memory_hooks(self):
Sylvain Gugger's avatar
Sylvain Gugger committed
943
944
945
        """
        Add a memory hook before and after each sub-module forward pass to record increase in memory consumption.

Sylvain Gugger's avatar
Sylvain Gugger committed
946
947
        Increase in memory consumption is stored in a `mem_rss_diff` attribute for each module and can be reset to zero
        with `model.reset_memory_hooks_state()`.
948
949
950
951
952
953
954
        """
        for module in self.modules():
            module.register_forward_pre_hook(self._hook_rss_memory_pre_forward)
            module.register_forward_hook(self._hook_rss_memory_post_forward)
        self.reset_memory_hooks_state()

    def reset_memory_hooks_state(self):
Sylvain Gugger's avatar
Sylvain Gugger committed
955
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
956
        Reset the `mem_rss_diff` attribute of each module (see [`~modeling_utils.ModuleUtilsMixin.add_memory_hooks`]).
Sylvain Gugger's avatar
Sylvain Gugger committed
957
        """
958
959
960
961
962
        for module in self.modules():
            module.mem_rss_diff = 0
            module.mem_rss_post_forward = 0
            module.mem_rss_pre_forward = 0

963
    @property
Sylvain Gugger's avatar
Sylvain Gugger committed
964
    def device(self) -> torch.device:
965
        """
966
        `torch.device`: The device on which the module is (assuming that all the module parameters are on the same
967
        device).
968
        """
Lysandre Debut's avatar
Lysandre Debut committed
969
        return get_parameter_device(self)
970

971
    @property
972
    def dtype(self) -> torch.dtype:
973
        """
974
        `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).
975
        """
Lysandre Debut's avatar
Lysandre Debut committed
976
        return get_parameter_dtype(self)
977
978

    def invert_attention_mask(self, encoder_attention_mask: Tensor) -> Tensor:
Sylvain Gugger's avatar
Sylvain Gugger committed
979
980
981
982
        """
        Invert an attention mask (e.g., switches 0. and 1.).

        Args:
983
            encoder_attention_mask (`torch.Tensor`): An attention mask.
Sylvain Gugger's avatar
Sylvain Gugger committed
984
985

        Returns:
986
            `torch.Tensor`: The inverted attention mask.
Sylvain Gugger's avatar
Sylvain Gugger committed
987
        """
988
989
990
991
992
993
994
995
996
997
        if encoder_attention_mask.dim() == 3:
            encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
        if encoder_attention_mask.dim() == 2:
            encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
        # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
        # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow
        # /transformer/transformer_layers.py#L270
        # encoder_extended_attention_mask = (encoder_extended_attention_mask ==
        # encoder_extended_attention_mask.transpose(-1, -2))
        encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
Yih-Dar's avatar
Yih-Dar committed
998
        encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * torch.finfo(self.dtype).min
999

1000
1001
        return encoder_extended_attention_mask

1002
    @staticmethod
1003
1004
1005
1006
1007
1008
1009
    def create_extended_attention_mask_for_decoder(input_shape, attention_mask, device=None):
        if device is not None:
            warnings.warn(
                "The `device` argument is deprecated and will be removed in v5 of Transformers.", FutureWarning
            )
        else:
            device = attention_mask.device
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
        batch_size, seq_length = input_shape
        seq_ids = torch.arange(seq_length, device=device)
        causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
        # in case past_key_values are used we need to add a prefix ones mask to the causal mask
        # causal and attention masks must have same type with pytorch version < 1.3
        causal_mask = causal_mask.to(attention_mask.dtype)

        if causal_mask.shape[1] < attention_mask.shape[1]:
            prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
            causal_mask = torch.cat(
                [
                    torch.ones((batch_size, seq_length, prefix_seq_len), device=device, dtype=causal_mask.dtype),
                    causal_mask,
                ],
                axis=-1,
            )

        extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
        return extended_attention_mask

1030
    def get_extended_attention_mask(
1031
        self, attention_mask: Tensor, input_shape: Tuple[int], device: torch.device = None, dtype: torch.float = None
1032
    ) -> Tensor:
Sylvain Gugger's avatar
Sylvain Gugger committed
1033
1034
        """
        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
1035
1036

        Arguments:
1037
            attention_mask (`torch.Tensor`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1038
                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
1039
            input_shape (`Tuple[int]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1040
                The shape of the input to the model.
1041
1042

        Returns:
1043
            `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
1044
        """
Yih-Dar's avatar
Yih-Dar committed
1045
1046
1047
        if dtype is None:
            dtype = self.dtype

1048
1049
1050
1051
1052
1053
        if not (attention_mask.dim() == 2 and self.config.is_decoder):
            # show warning only if it won't be shown in `create_extended_attention_mask_for_decoder`
            if device is not None:
                warnings.warn(
                    "The `device` argument is deprecated and will be removed in v5 of Transformers.", FutureWarning
                )
1054
1055
1056
1057
1058
1059
1060
1061
1062
        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
        # ourselves in which case we just need to make it broadcastable to all heads.
        if attention_mask.dim() == 3:
            extended_attention_mask = attention_mask[:, None, :, :]
        elif attention_mask.dim() == 2:
            # Provided a padding mask of dimensions [batch_size, seq_length]
            # - if the model is a decoder, apply a causal mask in addition to the padding mask
            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
            if self.config.is_decoder:
1063
                extended_attention_mask = ModuleUtilsMixin.create_extended_attention_mask_for_decoder(
1064
1065
                    input_shape, attention_mask, device
                )
1066
1067
1068
1069
            else:
                extended_attention_mask = attention_mask[:, None, None, :]
        else:
            raise ValueError(
1070
                f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"
1071
1072
1073
1074
            )

        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
        # masked positions, this operation will create a tensor which is 0.0 for
1075
        # positions we want to attend and the dtype's smallest value for masked positions.
1076
1077
        # Since we are adding it to the raw scores before the softmax, this is
        # effectively the same as removing these entirely.
Yih-Dar's avatar
Yih-Dar committed
1078
1079
        extended_attention_mask = extended_attention_mask.to(dtype=dtype)  # fp16 compatibility
        extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(dtype).min
1080
1081
        return extended_attention_mask

Sylvain Gugger's avatar
Sylvain Gugger committed
1082
1083
1084
    def get_head_mask(
        self, head_mask: Optional[Tensor], num_hidden_layers: int, is_attention_chunked: bool = False
    ) -> Tensor:
1085
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
1086
1087
1088
        Prepare the head mask if needed.

        Args:
1089
            head_mask (`torch.Tensor` with shape `[num_heads]` or `[num_hidden_layers x num_heads]`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
1090
                The mask indicating if we should keep the heads or not (1.0 for keep, 0.0 for discard).
1091
            num_hidden_layers (`int`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1092
                The number of hidden layers in the model.
1093
            is_attention_chunked (`bool`, *optional*, defaults to `False`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1094
1095
                Whether or not the attentions scores are computed by chunks or not.

1096
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
1097
1098
            `torch.Tensor` with shape `[num_hidden_layers x batch x num_heads x seq_length x seq_length]` or list with
            `[None]` for each layer.
1099
1100
1101
        """
        if head_mask is not None:
            head_mask = self._convert_head_mask_to_5d(head_mask, num_hidden_layers)
Patrick von Platen's avatar
Patrick von Platen committed
1102
1103
            if is_attention_chunked is True:
                head_mask = head_mask.unsqueeze(-1)
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
        else:
            head_mask = [None] * num_hidden_layers

        return head_mask

    def _convert_head_mask_to_5d(self, head_mask, num_hidden_layers):
        """-> [num_hidden_layers x batch x num_heads x seq_length x seq_length]"""
        if head_mask.dim() == 1:
            head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
            head_mask = head_mask.expand(num_hidden_layers, -1, -1, -1, -1)
        elif head_mask.dim() == 2:
            head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
        assert head_mask.dim() == 5, f"head_mask.dim != 5, instead {head_mask.dim()}"
1117
        head_mask = head_mask.to(dtype=self.dtype)  # switch to float if need + fp16 compatibility
1118
1119
        return head_mask

1120
1121
1122
1123
1124
    def num_parameters(self, only_trainable: bool = False, exclude_embeddings: bool = False) -> int:
        """
        Get number of (optionally, trainable or non-embeddings) parameters in the module.

        Args:
1125
            only_trainable (`bool`, *optional*, defaults to `False`):
1126
1127
                Whether or not to return only the number of trainable parameters

1128
            exclude_embeddings (`bool`, *optional*, defaults to `False`):
1129
1130
1131
                Whether or not to return only the number of non-embeddings parameters

        Returns:
1132
            `int`: The number of parameters.
1133
1134
        """

1135
1136
1137
1138
        if exclude_embeddings:
            embedding_param_names = [
                f"{name}.weight" for name, module_type in self.named_modules() if isinstance(module_type, nn.Embedding)
            ]
1139
            total_parameters = [
1140
1141
1142
                parameter for name, parameter in self.named_parameters() if name not in embedding_param_names
            ]
        else:
1143
1144
1145
1146
            total_parameters = list(self.parameters())

        total_numel = []
        is_loaded_in_4bit = getattr(self, "is_loaded_in_4bit", False)
1147

1148
1149
1150
1151
1152
1153
        if is_loaded_in_4bit:
            if is_bitsandbytes_available():
                import bitsandbytes as bnb
            else:
                raise ValueError(
                    "bitsandbytes is not installed but it seems that the model has been loaded in 4bit precision, something went wrong"
1154
                    " make sure to install bitsandbytes with `pip install bitsandbytes`. You also need a GPU. "
1155
1156
1157
1158
1159
1160
1161
                )

        for param in total_parameters:
            if param.requires_grad or not only_trainable:
                # For 4bit models, we need to multiply the number of parameters by 2 as half of the parameters are
                # used for the 4bit quantization (uint8 tensors are stored)
                if is_loaded_in_4bit and isinstance(param, bnb.nn.Params4bit):
1162
1163
1164
1165
                    quant_storage = self.hf_quantizer.quantization_config.bnb_4bit_quant_storage
                    # For compatibility with older PT version - see: https://github.com/huggingface/peft/pull/1635
                    nb_params = (
                        quant_storage.itemsize if hasattr(quant_storage, "itemsize") else quant_storage.element_size()
1166
                    )
1167
                    total_numel.append(param.numel() * 2 * nb_params)
1168
1169
1170
1171
                else:
                    total_numel.append(param.numel())

        return sum(total_numel)
1172
1173
1174
1175
1176
1177

    def estimate_tokens(self, input_dict: Dict[str, Union[torch.Tensor, Any]]) -> int:
        """
        Helper function to estimate the total number of tokens from the model inputs.

        Args:
1178
            inputs (`dict`): The model inputs.
1179
1180

        Returns:
1181
            `int`: The total number of tokens.
1182
        """
1183
1184
        if not hasattr(self, "warnings_issued"):
            self.warnings_issued = {}
1185
1186
        if self.main_input_name in input_dict:
            return input_dict[self.main_input_name].numel()
1187
        elif "estimate_tokens" not in self.warnings_issued:
1188
            logger.warning(
1189
1190
                "Could not estimate the number of tokens of the input, floating-point operations will not be computed"
            )
1191
1192
            self.warnings_issued["estimate_tokens"] = True
        return 0
1193
1194
1195
1196
1197
1198
1199

    def floating_point_ops(
        self, input_dict: Dict[str, Union[torch.Tensor, Any]], exclude_embeddings: bool = True
    ) -> int:
        """
        Get number of (optionally, non-embeddings) floating-point operations for the forward and backward passes of a
        batch with this transformer model. Default approximation neglects the quadratic dependency on the number of
Sylvain Gugger's avatar
Sylvain Gugger committed
1200
1201
        tokens (valid if `12 * d_model << sequence_length`) as laid out in [this
        paper](https://arxiv.org/pdf/2001.08361.pdf) section 2.1. Should be overridden for transformers with parameter
Sylvain Gugger's avatar
Sylvain Gugger committed
1202
        re-use e.g. Albert or Universal Transformers, or if doing long-range modeling with very high sequence lengths.
1203
1204

        Args:
1205
            batch_size (`int`):
1206
1207
                The batch size for the forward pass.

1208
            sequence_length (`int`):
1209
1210
                The number of tokens in each line of the batch.

1211
            exclude_embeddings (`bool`, *optional*, defaults to `True`):
1212
1213
1214
                Whether or not to count embedding and softmax operations.

        Returns:
1215
            `int`: The number of floating-point operations.
1216
1217
1218
1219
        """

        return 6 * self.estimate_tokens(input_dict) * self.num_parameters(exclude_embeddings=exclude_embeddings)

Julien Chaumond's avatar
Julien Chaumond committed
1220

1221
class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin, PeftAdapterMixin):
1222
1223
    r"""
    Base class for all models.
1224

Sylvain Gugger's avatar
Sylvain Gugger committed
1225
1226
    [`PreTrainedModel`] takes care of storing the configuration of the models and handles methods for loading,
    downloading and saving models as well as a few methods common to all models to:
1227

1228
1229
        - resize the input embeddings,
        - prune heads in the self-attention heads.
1230

1231
    Class attributes (overridden by derived classes):
Sylvain Gugger's avatar
Sylvain Gugger committed
1232

Sylvain Gugger's avatar
Sylvain Gugger committed
1233
1234
1235
1236
        - **config_class** ([`PretrainedConfig`]) -- A subclass of [`PretrainedConfig`] to use as configuration class
          for this model architecture.
        - **load_tf_weights** (`Callable`) -- A python *method* for loading a TensorFlow checkpoint in a PyTorch model,
          taking as arguments:
1237

Sylvain Gugger's avatar
Sylvain Gugger committed
1238
1239
            - **model** ([`PreTrainedModel`]) -- An instance of the model on which to load the TensorFlow checkpoint.
            - **config** ([`PreTrainedConfig`]) -- An instance of the configuration associated to the model.
1240
            - **path** (`str`) -- A path to the TensorFlow checkpoint.
1241

Sylvain Gugger's avatar
Sylvain Gugger committed
1242
1243
        - **base_model_prefix** (`str`) -- A string indicating the attribute associated to the base model in derived
          classes of the same architecture adding modules on top of the base model.
1244
        - **is_parallelizable** (`bool`) -- A flag indicating whether this model supports model parallelization.
Sylvain Gugger's avatar
Sylvain Gugger committed
1245
1246
        - **main_input_name** (`str`) -- The name of the principal input to the model (often `input_ids` for NLP
          models, `pixel_values` for vision models and `input_values` for speech models).
1247
    """
1248

1249
    config_class = None
1250
    base_model_prefix = ""
1251
    main_input_name = "input_ids"
1252
1253
    model_tags = None

1254
    _auto_class = None
1255
    _no_split_modules = None
1256
    _skip_keys_device_placement = None
1257
    _keep_in_fp32_modules = None
1258

1259
1260
    # a list of `re` patterns of `state_dict` keys that should be removed from the list of missing
    # keys we find (keys inside the model but not in the checkpoint) and avoid unnecessary warnings.
1261
    _keys_to_ignore_on_load_missing = None
1262
1263
1264
    # a list of `re` patterns of `state_dict` keys that should be removed from the list of
    # unexpected keys we find (keys inside the checkpoint but not the model) and avoid unnecessary
    # warnings.
1265
    _keys_to_ignore_on_load_unexpected = None
1266
1267
    # a list of `state_dict` keys to ignore when saving the model (useful for keys that aren't
    # trained, but which are either deterministic or tied variables)
1268
    _keys_to_ignore_on_save = None
Sylvain Gugger's avatar
Sylvain Gugger committed
1269
1270
    # a list of `state_dict` keys that are potentially tied to another key in the state_dict.
    _tied_weights_keys = None
1271

1272
    is_parallelizable = False
1273
    supports_gradient_checkpointing = False
1274

1275
1276
1277
    # Flash Attention 2 support
    _supports_flash_attn_2 = False

1278
1279
1280
    # SDPA support
    _supports_sdpa = False

1281
1282
1283
    # Has support for a `Cache` instance as `past_key_values`
    _supports_cache_class = False

1284
    @property
1285
    def dummy_inputs(self) -> Dict[str, torch.Tensor]:
1286
        """
1287
        `Dict[str, torch.Tensor]`: Dummy inputs to do a forward pass in the network.
1288
        """
1289
        return {"input_ids": torch.tensor(DUMMY_INPUTS)}
1290

1291
1292
1293
1294
1295
1296
1297
    @property
    def framework(self) -> str:
        """
        :str: Identifies that this is a PyTorch model.
        """
        return "pt"

1298
    def __init__(self, config: PretrainedConfig, *inputs, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
1299
        super().__init__()
1300
1301
        if not isinstance(config, PretrainedConfig):
            raise ValueError(
1302
1303
1304
                f"Parameter config in `{self.__class__.__name__}(config)` should be an instance of class "
                "`PretrainedConfig`. To create a model from a pretrained model use "
                f"`model = {self.__class__.__name__}.from_pretrained(PRETRAINED_MODEL_NAME)`"
1305
            )
1306
        # Save config and origin of the pretrained weights if given in model
1307
1308
1309
        config = self._autoset_attn_implementation(
            config, torch_dtype=torch.get_default_dtype(), check_device_map=False
        )
1310
        self.config = config
1311

1312
        self.name_or_path = config.name_or_path
1313
        self.warnings_issued = {}
1314
        self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
1315
1316
1317
1318
        # Overwrite the class attribute to make it an instance attribute, so models like
        # `InstructBlipForConditionalGeneration` can dynamically update it without modifying the class attribute
        # when a different component (e.g. language_model) is used.
        self._keep_in_fp32_modules = copy.copy(self.__class__._keep_in_fp32_modules)
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332

    def post_init(self):
        """
        A method executed at the end of each Transformer model initialization, to execute code that needs the model's
        modules properly initialized (such as weight initialization).
        """
        self.init_weights()
        self._backward_compatibility_gradient_checkpointing()

    def _backward_compatibility_gradient_checkpointing(self):
        if self.supports_gradient_checkpointing and getattr(self.config, "gradient_checkpointing", False):
            self.gradient_checkpointing_enable()
            # Remove the attribute now that is has been consumed, so it's no saved in the config.
            delattr(self.config, "gradient_checkpointing")
1333

1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
    def add_model_tags(self, tags: Union[List[str], str]) -> None:
        r"""
        Add custom tags into the model that gets pushed to the Hugging Face Hub. Will
        not overwrite existing tags in the model.

        Args:
            tags (`Union[List[str], str]`):
                The desired tags to inject in the model

        Examples:

        ```python
        from transformers import AutoModel

1348
        model = AutoModel.from_pretrained("google-bert/bert-base-cased")
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365

        model.add_model_tags(["custom", "custom-bert"])

        # Push the model to your namespace with the name "my-custom-bert".
        model.push_to_hub("my-custom-bert")
        ```
        """
        if isinstance(tags, str):
            tags = [tags]

        if self.model_tags is None:
            self.model_tags = []

        for tag in tags:
            if tag not in self.model_tags:
                self.model_tags.append(tag)

1366
1367
1368
1369
1370
1371
    @classmethod
    def _from_config(cls, config, **kwargs):
        """
        All context managers that the model should be initialized under go here.

        Args:
1372
1373
            torch_dtype (`torch.dtype`, *optional*):
                Override the default `torch.dtype` and load the model under this dtype.
1374
1375
        """
        torch_dtype = kwargs.pop("torch_dtype", None)
1376
        use_flash_attention_2 = kwargs.pop("use_flash_attention_2", False)
1377
1378
1379
1380
1381
1382

        # override default dtype if needed
        dtype_orig = None
        if torch_dtype is not None:
            dtype_orig = cls._set_default_torch_dtype(torch_dtype)

1383
1384
1385
        config = copy.deepcopy(config)  # We do not want to modify the config inplace in _from_config.
        config._attn_implementation = kwargs.pop("attn_implementation", None)
        config = cls._autoset_attn_implementation(
1386
1387
1388
1389
            config,
            use_flash_attention_2=use_flash_attention_2,
            check_device_map=False,
            torch_dtype=torch_dtype,
1390
        )
1391

1392
1393
1394
1395
1396
1397
        if is_deepspeed_zero3_enabled():
            import deepspeed

            logger.info("Detected DeepSpeed ZeRO-3: activating zero.init() for this model")
            # this immediately partitions the model across all gpus, to avoid the overhead in time
            # and memory copying it on CPU or each GPU first
1398
            with deepspeed.zero.Init(config_dict_or_path=deepspeed_config()):
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
                model = cls(config, **kwargs)
        else:
            model = cls(config, **kwargs)

        # restore default dtype if it was modified
        if dtype_orig is not None:
            torch.set_default_dtype(dtype_orig)

        return model

1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
    @classmethod
    def _autoset_attn_implementation(
        cls,
        config,
        use_flash_attention_2: bool = False,
        torch_dtype: Optional[torch.dtype] = None,
        device_map: Optional[Union[str, Dict[str, int]]] = None,
        check_device_map: bool = True,
    ):
        """
        Automatically checks and dispatches to a default attention implementation. In order of priority:
            1. An implementation specified in `config._attn_implementation` (due for example to the argument attn_implementation="sdpa" in from_pretrained).
            2. DEPRECATED: if use_flash_attention_2 is set to `True` and `flash_attn` is available, flash attention. (`LlamaFlashAttention` for example)
            3. SDPA implementation, if available and supported by the model type. (`LlamaSdpaAttention` for example)
            4. The default model's implementation otherwise (`LlamaAttention` for example) .
        """
        # Here we use config._attn_implementation_internal to check whether the attention implementation was explicitely set by the user.
        # The property `PretrainedConfig._attn_implementation` is never `None`, for backward compatibility (always fall back on "eager").
        # The `hasattr` here is used as some Transformers tests for some reason do not call PretrainedConfig __init__ (e.g. test_no_super_init_config_and_model)
1428
        requested_attn_implementation = None
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
        if hasattr(config, "_attn_implementation_internal") and config._attn_implementation_internal is not None:
            if config._attn_implementation != "flash_attention_2" and use_flash_attention_2:
                raise ValueError(
                    f'Both attn_implementation="{config._attn_implementation}" and `use_flash_attention_2=True` were used when loading the model, which are not compatible.'
                    ' We recommend to just use `attn_implementation="flash_attention_2"` when loading the model.'
                )

            if config._attn_implementation not in ["eager", "sdpa", "flash_attention_2"]:
                message = f'Specified `attn_implementation="{config._attn_implementation}"` is not supported. The only possible arguments are `attn_implementation="eager"` (manual attention implementation)'
                if cls._supports_flash_attn_2:
                    message += ', `"attn_implementation=flash_attention_2"` (implementation using flash attention 2)'
                if cls._supports_sdpa:
                    message += ', `"attn_implementation=sdpa"` (implementation using torch.nn.functional.scaled_dot_product_attention)'
                raise ValueError(message + ".")

            # If a config is passed with a preset attn_implementation, we skip the automatic dispatch and use the user-provided config, with hard checks that the requested attention implementation is available.
1445
            requested_attn_implementation = config._attn_implementation_internal
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457

        if use_flash_attention_2:
            logger.warning_once(
                'The model was loaded with use_flash_attention_2=True, which is deprecated and may be removed in a future release. Please use `attn_implementation="flash_attention_2"` instead.'
            )
            config._attn_implementation = "flash_attention_2"

        if config._attn_implementation == "flash_attention_2":
            cls._check_and_enable_flash_attn_2(
                config,
                torch_dtype=torch_dtype,
                device_map=device_map,
1458
                hard_check_only=False,
1459
1460
                check_device_map=check_device_map,
            )
1461
        elif requested_attn_implementation in [None, "sdpa"] and not is_torch_xla_available():
1462
            # use_flash_attention_2 takes priority over SDPA, hence SDPA treated in this elif.
1463
            config = cls._check_and_enable_sdpa(
1464
1465
                config,
                hard_check_only=False if requested_attn_implementation is None else True,
1466
1467
            )
        else:
1468
1469
1470
1471
            config._attn_implementation = "eager"

        return config

1472
1473
1474
1475
1476
1477
1478
    @classmethod
    def _set_default_torch_dtype(cls, dtype: torch.dtype) -> torch.dtype:
        """
        Change the default dtype and return the previous one. This is needed when wanting to instantiate the model
        under specific dtype.

        Args:
1479
            dtype (`torch.dtype`):
1480
1481
1482
                a floating dtype to set to.

        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
1483
1484
            `torch.dtype`: the original `dtype` that can be used to restore `torch.set_default_dtype(dtype)` if it was
            modified. If it wasn't, returns `None`.
1485

1486
1487
        Note `set_default_dtype` currently only works with floating-point types and asserts if for example,
        `torch.int64` is passed. So if a non-float `dtype` is passed this functions will throw an exception.
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
        """
        if not dtype.is_floating_point:
            raise ValueError(
                f"Can't instantiate {cls.__name__} model under dtype={dtype} since it is not a floating point dtype"
            )

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

1499
    @property
1500
1501
    def base_model(self) -> nn.Module:
        """
1502
        `torch.nn.Module`: The main body of the model.
1503
        """
1504
        return getattr(self, self.base_model_prefix, self)
thomwolf's avatar
thomwolf committed
1505

1506
1507
    @classmethod
    def can_generate(cls) -> bool:
1508
1509
1510
1511
1512
1513
        """
        Returns whether this model can generate sequences with `.generate()`.

        Returns:
            `bool`: Whether this model can generate sequences with `.generate()`.
        """
1514
1515
1516
        # Detects whether `prepare_inputs_for_generation` has been overwritten, which is a requirement for generation.
        # Alternativelly, the model can also have a custom `generate` function.
        if "GenerationMixin" in str(cls.prepare_inputs_for_generation) and "GenerationMixin" in str(cls.generate):
1517
1518
1519
            return False
        return True

1520
1521
    @classmethod
    def _check_and_enable_flash_attn_2(
1522
1523
1524
1525
1526
1527
        cls,
        config,
        torch_dtype: Optional[torch.dtype] = None,
        device_map: Optional[Union[str, Dict[str, int]]] = None,
        check_device_map: bool = True,
        hard_check_only: bool = False,
1528
1529
    ) -> PretrainedConfig:
        """
1530
        Checks the availability of Flash Attention 2 and compatibility with the current model.
1531

1532
        If all checks pass and `hard_check_only` is False, the method will set the config attribute `attn_implementation` to "flash_attention_2" so that the model can initialize the correct attention module.
1533
1534
1535
        """
        if not cls._supports_flash_attn_2:
            raise ValueError(
1536
1537
1538
                f"{cls.__name__} does not support Flash Attention 2.0 yet. Please request to add support where"
                f" the model is hosted, on its model hub page: https://huggingface.co/{config._name_or_path}/discussions/new"
                " or in the Transformers GitHub repo: https://github.com/huggingface/transformers/issues/new"
1539
1540
            )

1541
        if not is_flash_attn_2_available():
1542
1543
1544
            preface = "FlashAttention2 has been toggled on, but it cannot be used due to the following error:"
            install_message = "Please refer to the documentation of https://huggingface.co/docs/transformers/perf_infer_gpu_one#flashattention-2 to install Flash Attention 2."

1545
1546
1547
1548
1549
            if importlib.util.find_spec("flash_attn") is None:
                raise ImportError(f"{preface} the package flash_attn seems to be not installed. {install_message}")

            flash_attention_version = version.parse(importlib.metadata.version("flash_attn"))
            if torch.version.cuda:
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
                if flash_attention_version < version.parse("2.1.0"):
                    raise ImportError(
                        f"{preface} you need flash_attn package version to be greater or equal than 2.1.0. Detected version {flash_attention_version}. {install_message}"
                    )
                else:
                    raise ImportError(f"{preface} Flash Attention 2 is not available. {install_message}")
            elif torch.version.hip:
                if flash_attention_version < version.parse("2.0.4"):
                    raise ImportError(
                        f"{preface} you need flash_attn package version to be greater or equal than 2.0.4. Make sure to have that version installed - detected version {flash_attention_version}. {install_message}"
                    )
                else:
                    raise ImportError(f"{preface} Flash Attention 2 is not available. {install_message}")
1563
1564
1565
1566
1567
1568
1569
1570
1571

        _is_bettertransformer = getattr(cls, "use_bettertransformer", False)

        if _is_bettertransformer:
            raise ValueError(
                "Flash Attention 2 and BetterTransformer API are not compatible. Please make sure to disable BetterTransformers by doing model.reverse_bettertransformer()"
            )

        if torch_dtype is None:
1572
            logger.warning_once(
1573
1574
1575
                "You are attempting to use Flash Attention 2.0 without specifying a torch dtype. This might lead to unexpected behaviour"
            )
        elif torch_dtype is not None and torch_dtype not in [torch.float16, torch.bfloat16]:
1576
1577
1578
1579
            logger.warning_once(
                "Flash Attention 2.0 only supports torch.float16 and torch.bfloat16 dtypes, but"
                f" the current dype in {cls.__name__} is {torch_dtype}. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator,"
                ' or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`'
1580
1581
            )

1582
1583
1584
        # The check `torch.empty(0).device.type != "cuda"` is needed as the model may be initialized after `torch.set_default_device` has been called,
        # or the model may be initialized under the context manager `with torch.device("cuda"):`.
        if check_device_map and device_map is None and torch.empty(0).device.type != "cuda":
1585
            if torch.cuda.is_available():
1586
                logger.warning_once(
1587
                    "You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU"
1588
1589
1590
1591
                    " after initializing it on CPU with `model.to('cuda')`."
                )
            else:
                raise ValueError(
1592
                    "You are attempting to use Flash Attention 2.0 with a model not initialized on GPU and with no GPU available. "
1593
1594
1595
1596
                    "This is not supported yet. Please make sure to have access to a GPU and either initialise the model on a GPU by passing a device_map "
                    "or initialising the model on CPU and then moving it to GPU."
                )
        elif (
1597
1598
            check_device_map
            and device_map is not None
1599
1600
1601
1602
1603
1604
1605
            and isinstance(device_map, dict)
            and ("cpu" in device_map.values() or "disk" in device_map.values())
        ):
            raise ValueError(
                "You are attempting to use Flash Attention 2.0 with a model dispatched on CPU or disk. This is not supported. Please make sure to "
                "initialise the model on a GPU by passing a device_map that contains only GPU devices as keys."
            )
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
        if not hard_check_only:
            config._attn_implementation = "flash_attention_2"
        return config

    @classmethod
    def _check_and_enable_sdpa(cls, config, hard_check_only: bool = False) -> PretrainedConfig:
        """
        Checks the availability of SDPA for a given model.

        If all checks pass and `hard_check_only` is False, the method will set the config attribute `_attn_implementation` to "flash_attention_2" so that the model can initialize the correct attention module.
        """
        if hard_check_only:
            if not cls._supports_sdpa:
                raise ValueError(
1620
1621
1622
                    f"{cls.__name__} does not support an attention implementation through torch.nn.functional.scaled_dot_product_attention yet."
                    " Please request the support for this architecture: https://github.com/huggingface/transformers/issues/28005. If you believe"
                    ' this error is a bug, please open an issue in Transformers GitHub repository and load your model with the argument `attn_implementation="eager"` meanwhile. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="eager")`'
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
                )
            if not is_torch_sdpa_available():
                raise ImportError(
                    "PyTorch SDPA requirements in Transformers are not met. Please install torch>=2.1.1."
                )

        if not is_torch_sdpa_available() or not cls._supports_sdpa:
            return config

        _is_bettertransformer = getattr(cls, "use_bettertransformer", False)
        if _is_bettertransformer:
            return config

        if not hard_check_only:
            config._attn_implementation = "sdpa"
1638
1639
        return config

1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
    def enable_input_require_grads(self):
        """
        Enables the gradients for the input embeddings. This is useful for fine-tuning adapter weights while keeping
        the model weights fixed.
        """

        def make_inputs_require_grads(module, input, output):
            output.requires_grad_(True)

        self._require_grads_hook = self.get_input_embeddings().register_forward_hook(make_inputs_require_grads)

    def disable_input_require_grads(self):
        """
        Removes the `_require_grads_hook`.
        """
        self._require_grads_hook.remove()

1657
    def get_input_embeddings(self) -> nn.Module:
1658
1659
1660
1661
        """
        Returns the model's input embeddings.

        Returns:
1662
            `nn.Module`: A torch module mapping vocabulary to hidden states.
thomwolf's avatar
thomwolf committed
1663
        """
1664
        base_model = getattr(self, self.base_model_prefix, self)
thomwolf's avatar
thomwolf committed
1665
1666
1667
1668
        if base_model is not self:
            return base_model.get_input_embeddings()
        else:
            raise NotImplementedError
thomwolf's avatar
thomwolf committed
1669

1670
    def set_input_embeddings(self, value: nn.Module):
1671
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
1672
        Set model's input embeddings.
1673
1674

        Args:
1675
            value (`nn.Module`): A module mapping vocabulary to hidden states.
thomwolf's avatar
thomwolf committed
1676
1677
1678
1679
1680
1681
        """
        base_model = getattr(self, self.base_model_prefix, self)
        if base_model is not self:
            base_model.set_input_embeddings(value)
        else:
            raise NotImplementedError
thomwolf's avatar
thomwolf committed
1682

1683
    def get_output_embeddings(self) -> nn.Module:
1684
1685
1686
1687
        """
        Returns the model's output embeddings.

        Returns:
1688
            `nn.Module`: A torch module mapping hidden states to vocabulary.
thomwolf's avatar
thomwolf committed
1689
        """
1690
        return None  # Overwrite for models with output embeddings
thomwolf's avatar
thomwolf committed
1691

1692
1693
    def _init_weights(self, module):
        """
1694
1695
1696
1697
        Initialize the weights. This method should be overridden by derived class and is
        the only initialization method that will be called when loading a checkpoint
        using `from_pretrained`. Any attempt to initialize outside of this function
        will be useless as the torch.nn.init function are all replaced with skip.
1698
        """
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
        pass

    def _initialize_weights(self, module):
        """
        Initialize the weights if they are not already initialized.
        """
        if getattr(module, "_is_hf_initialized", False):
            return
        self._init_weights(module)
        module._is_hf_initialized = True
1709

1710
    def tie_weights(self):
1711
1712
        """
        Tie the weights between the input embeddings and the output embeddings.
1713

Sylvain Gugger's avatar
Sylvain Gugger committed
1714
1715
        If the `torchscript` flag is set in the configuration, can't handle parameter sharing so we are cloning the
        weights instead.
thomwolf's avatar
thomwolf committed
1716
        """
1717
1718
1719
1720
        if getattr(self.config, "tie_word_embeddings", True):
            output_embeddings = self.get_output_embeddings()
            if output_embeddings is not None:
                self._tie_or_clone_weights(output_embeddings, self.get_input_embeddings())
thomwolf's avatar
thomwolf committed
1721

1722
        if getattr(self.config, "is_encoder_decoder", False) and getattr(self.config, "tie_encoder_decoder", False):
Weizhen's avatar
Weizhen committed
1723
1724
            if hasattr(self, self.base_model_prefix):
                self = getattr(self, self.base_model_prefix)
1725
1726
1727
1728
1729
1730
1731
            tied_weights = self._tie_encoder_decoder_weights(
                self.encoder, self.decoder, self.base_model_prefix, "encoder"
            )
            # Setting a dynamic variable instead of `_tied_weights_keys` because it's a class
            # attributed not an instance member, therefore modifying it will modify the entire class
            # Leading to issues on subsequent calls by different tests or subsequent calls.
            self._dynamic_tied_weights_keys = tied_weights
1732

Sylvain Gugger's avatar
Sylvain Gugger committed
1733
1734
1735
1736
        for module in self.modules():
            if hasattr(module, "_tie_weights"):
                module._tie_weights()

1737
    @staticmethod
1738
1739
1740
    def _tie_encoder_decoder_weights(
        encoder: nn.Module, decoder: nn.Module, base_model_prefix: str, base_encoder_name: str
    ):
1741
        uninitialized_encoder_weights: List[str] = []
1742
        tied_weights: List[str] = []
Weizhen's avatar
Weizhen committed
1743
1744
        if decoder.__class__ != encoder.__class__:
            logger.info(
Sylvain Gugger's avatar
Sylvain Gugger committed
1745
1746
                f"{decoder.__class__} and {encoder.__class__} are not equal. In this case make sure that all encoder"
                " weights are correctly initialized."
Weizhen's avatar
Weizhen committed
1747
            )
1748
1749
1750
1751
1752

        def tie_encoder_to_decoder_recursively(
            decoder_pointer: nn.Module,
            encoder_pointer: nn.Module,
            module_name: str,
1753
            base_encoder_name: str,
1754
1755
            uninitialized_encoder_weights: List[str],
            depth=0,
1756
1757
            total_decoder_name="",
            total_encoder_name="",
1758
1759
1760
        ):
            assert isinstance(decoder_pointer, nn.Module) and isinstance(
                encoder_pointer, nn.Module
1761
            ), f"{decoder_pointer} and {encoder_pointer} have to be of type nn.Module"
1762
1763
1764
            if hasattr(decoder_pointer, "weight"):
                assert hasattr(encoder_pointer, "weight")
                encoder_pointer.weight = decoder_pointer.weight
1765
                tied_weights.append(f"{base_encoder_name}{total_encoder_name}.weight")
1766
1767
                if hasattr(decoder_pointer, "bias"):
                    assert hasattr(encoder_pointer, "bias")
1768
                    tied_weights.append(f"{base_encoder_name}{total_encoder_name}.bias")
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
                    encoder_pointer.bias = decoder_pointer.bias
                return

            encoder_modules = encoder_pointer._modules
            decoder_modules = decoder_pointer._modules
            if len(decoder_modules) > 0:
                assert (
                    len(encoder_modules) > 0
                ), f"Encoder module {encoder_pointer} does not match decoder module {decoder_pointer}"

1779
                all_encoder_weights = {module_name + "/" + sub_name for sub_name in encoder_modules.keys()}
1780
1781
1782
1783
1784
                encoder_layer_pos = 0
                for name, module in decoder_modules.items():
                    if name.isdigit():
                        encoder_name = str(int(name) + encoder_layer_pos)
                        decoder_name = name
Weizhen's avatar
Weizhen committed
1785
1786
1787
                        if not isinstance(decoder_modules[decoder_name], type(encoder_modules[encoder_name])) and len(
                            encoder_modules
                        ) != len(decoder_modules):
1788
1789
                            # this can happen if the name corresponds to the position in a list module list of layers
                            # in this case the decoder has added a cross-attention that the encoder does not have
1790
                            # thus skip this step and subtract one layer pos from encoder
1791
1792
1793
1794
1795
1796
                            encoder_layer_pos -= 1
                            continue
                    elif name not in encoder_modules:
                        continue
                    elif depth > 500:
                        raise ValueError(
Sylvain Gugger's avatar
Sylvain Gugger committed
1797
1798
                            "Max depth of recursive function `tie_encoder_to_decoder` reached. It seems that there is"
                            " a circular dependency between two or more `nn.Modules` of your model."
1799
1800
1801
1802
1803
1804
1805
                        )
                    else:
                        decoder_name = encoder_name = name
                    tie_encoder_to_decoder_recursively(
                        decoder_modules[decoder_name],
                        encoder_modules[encoder_name],
                        module_name + "/" + name,
1806
                        base_encoder_name,
1807
1808
                        uninitialized_encoder_weights,
                        depth=depth + 1,
1809
1810
                        total_encoder_name=f"{total_encoder_name}.{encoder_name}",
                        total_decoder_name=f"{total_decoder_name}.{decoder_name}",
1811
1812
1813
1814
1815
1816
                    )
                    all_encoder_weights.remove(module_name + "/" + encoder_name)

                uninitialized_encoder_weights += list(all_encoder_weights)

        # tie weights recursively
1817
1818
1819
1820
        tie_encoder_to_decoder_recursively(
            decoder, encoder, base_model_prefix, base_encoder_name, uninitialized_encoder_weights
        )

1821
1822
1823
1824
        if len(uninitialized_encoder_weights) > 0:
            logger.warning(
                f"The following encoder weights were not tied to the decoder {uninitialized_encoder_weights}"
            )
1825
        return tied_weights
1826

1827
    def _tie_or_clone_weights(self, output_embeddings, input_embeddings):
Lysandre's avatar
Lysandre committed
1828
        """Tie or clone module weights depending of whether we are using TorchScript or not"""
thomwolf's avatar
thomwolf committed
1829
        if self.config.torchscript:
1830
            output_embeddings.weight = nn.Parameter(input_embeddings.weight.clone())
thomwolf's avatar
thomwolf committed
1831
        else:
1832
            output_embeddings.weight = input_embeddings.weight
thomwolf's avatar
thomwolf committed
1833

Sam Shleifer's avatar
Sam Shleifer committed
1834
        if getattr(output_embeddings, "bias", None) is not None:
1835
            output_embeddings.bias.data = nn.functional.pad(
1836
                output_embeddings.bias.data,
Lysandre's avatar
Lysandre committed
1837
1838
1839
1840
                (
                    0,
                    output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0],
                ),
1841
1842
                "constant",
                0,
1843
            )
1844
        if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
1845
            output_embeddings.out_features = input_embeddings.num_embeddings
1846

Marc Sun's avatar
Marc Sun committed
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
    def _get_no_split_modules(self, device_map: str):
        """
        Get the modules of the model that should not be spit 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
        """
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
        _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, PreTrainedModel):
                    if module._no_split_modules is None:
                        raise ValueError(
                            f"{module.__class__.__name__} does not support `device_map='{device_map}'`. To implement support, the model "
                            "class needs to implement the `_no_split_modules` attribute."
                        )
                    else:
                        _no_split_modules = _no_split_modules | set(module._no_split_modules)
                modules_to_check += list(module.children())
Marc Sun's avatar
Marc Sun committed
1874
1875
        return list(_no_split_modules)

1876
1877
1878
    def resize_token_embeddings(
        self, new_num_tokens: Optional[int] = None, pad_to_multiple_of: Optional[int] = None
    ) -> nn.Embedding:
1879
        """
1880
        Resizes input token embeddings matrix of the model if `new_num_tokens != config.vocab_size`.
1881

1882
        Takes care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
thomwolf's avatar
thomwolf committed
1883

1884
        Arguments:
1885
            new_num_tokens (`int`, *optional*):
1886
                The new number of tokens in the embedding matrix. Increasing the size will add newly initialized
Sylvain Gugger's avatar
Sylvain Gugger committed
1887
1888
                vectors at the end. Reducing the size will remove vectors from the end. If not provided or `None`, just
                returns a pointer to the input tokens `torch.nn.Embedding` module of the model without doing anything.
1889
            pad_to_multiple_of (`int`, *optional*):
Arthur's avatar
Arthur committed
1890
1891
                If set will pad the embedding matrix to a multiple of the provided value.If `new_num_tokens` is set to
                `None` will just pad the embedding to a multiple of `pad_to_multiple_of`.
1892
1893
1894
1895
1896

                This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability
                `>= 7.5` (Volta), or on TPUs which benefit from having sequence lengths be a multiple of 128. For more
                details about this, or help on choosing the correct value for resizing, refer to this guide:
                https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html#requirements-tc
1897
1898

        Return:
1899
            `torch.nn.Embedding`: Pointer to the input tokens Embeddings Module of the model.
thomwolf's avatar
thomwolf committed
1900
        """
1901
        model_embeds = self._resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
Arthur's avatar
Arthur committed
1902
        if new_num_tokens is None and pad_to_multiple_of is None:
thomwolf's avatar
thomwolf committed
1903
            return model_embeds
thomwolf's avatar
thomwolf committed
1904
1905

        # Update base model and current model config
Arthur's avatar
Arthur committed
1906
1907
        self.config.vocab_size = model_embeds.weight.shape[0]
        self.vocab_size = model_embeds.weight.shape[0]
thomwolf's avatar
thomwolf committed
1908
1909

        # Tie weights again if needed
1910
        self.tie_weights()
thomwolf's avatar
thomwolf committed
1911

thomwolf's avatar
thomwolf committed
1912
1913
        return model_embeds

1914
    def _resize_token_embeddings(self, new_num_tokens, pad_to_multiple_of=None):
thomwolf's avatar
thomwolf committed
1915
        old_embeddings = self.get_input_embeddings()
1916
        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens, pad_to_multiple_of)
1917
1918
1919
        if hasattr(old_embeddings, "_hf_hook"):
            hook = old_embeddings._hf_hook
            add_hook_to_module(new_embeddings, hook)
1920
1921
        old_embeddings_requires_grad = old_embeddings.weight.requires_grad
        new_embeddings.requires_grad_(old_embeddings_requires_grad)
thomwolf's avatar
thomwolf committed
1922
        self.set_input_embeddings(new_embeddings)
1923
        is_quantized = hasattr(self, "hf_quantizer") and self.hf_quantizer is not None
1924

1925
1926
        # Update new_num_tokens with the actual size of new_embeddings
        if pad_to_multiple_of is not None:
1927
            if is_deepspeed_zero3_enabled() and not is_quantized:
1928
1929
1930
1931
1932
1933
1934
                import deepspeed

                with deepspeed.zero.GatheredParameters(new_embeddings.weight, modifier_rank=None):
                    new_num_tokens = new_embeddings.weight.shape[0]
            else:
                new_num_tokens = new_embeddings.weight.shape[0]

1935
1936
1937
        # if word embeddings are not tied, make sure that lm head is resized as well
        if self.get_output_embeddings() is not None and not self.config.tie_word_embeddings:
            old_lm_head = self.get_output_embeddings()
1938
1939
1940
1941
            if isinstance(old_lm_head, torch.nn.Embedding):
                new_lm_head = self._get_resized_embeddings(old_lm_head, new_num_tokens)
            else:
                new_lm_head = self._get_resized_lm_head(old_lm_head, new_num_tokens)
1942
1943
1944
            if hasattr(old_lm_head, "_hf_hook"):
                hook = old_lm_head._hf_hook
                add_hook_to_module(new_lm_head, hook)
1945
1946
            old_lm_head_requires_grad = old_lm_head.weight.requires_grad
            new_lm_head.requires_grad_(old_lm_head_requires_grad)
1947
1948
            self.set_output_embeddings(new_lm_head)

thomwolf's avatar
thomwolf committed
1949
        return self.get_input_embeddings()
1950

1951
    def _get_resized_embeddings(
1952
1953
1954
1955
        self,
        old_embeddings: nn.Embedding,
        new_num_tokens: Optional[int] = None,
        pad_to_multiple_of: Optional[int] = None,
1956
    ) -> nn.Embedding:
1957
1958
1959
        """
        Build a resized Embedding Module from a provided token Embedding Module. Increasing the size will add newly
        initialized vectors at the end. Reducing the size will remove vectors from the end
1960
1961

        Args:
1962
            old_embeddings (`torch.nn.Embedding`):
1963
                Old embeddings to be resized.
1964
            new_num_tokens (`int`, *optional*):
1965
                New number of tokens in the embedding matrix.
1966
1967

                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
1968
                vectors from the end. If not provided or `None`, just returns a pointer to the input tokens
1969
                `torch.nn.Embedding` module of the model without doing anything.
1970
            pad_to_multiple_of (`int`, *optional*):
Arthur's avatar
Arthur committed
1971
1972
                If set will pad the embedding matrix to a multiple of the provided value. If `new_num_tokens` is set to
                `None` will just pad the embedding to a multiple of `pad_to_multiple_of`.
1973
1974
1975
1976
1977
1978

                This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability
                `>= 7.5` (Volta), or on TPUs which benefit from having sequence lengths be a multiple of 128. For more
                details about this, or help on choosing the correct value for resizing, refer to this guide:
                https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html#requirements-tc

1979
1980

        Return:
1981
1982
            `torch.nn.Embedding`: Pointer to the resized Embedding Module or the old Embedding Module if
            `new_num_tokens` is `None`
1983
        """
1984
1985
1986
1987
1988
1989
1990
1991

        if pad_to_multiple_of is not None:
            if not isinstance(pad_to_multiple_of, int):
                raise ValueError(
                    f"Asking to pad the embedding matrix to a multiple of `{pad_to_multiple_of}`, which is not and integer. Please make sure to pass an integer"
                )
            if new_num_tokens is None:
                new_num_tokens = old_embeddings.weight.shape[0]
1992
            new_num_tokens = ((new_num_tokens + pad_to_multiple_of - 1) // pad_to_multiple_of) * pad_to_multiple_of
1993
        else:
1994
            logger.info(
1995
                "You are resizing the embedding layer without providing a `pad_to_multiple_of` parameter. This means that the new embedding"
1996
                f" dimension will be {new_num_tokens}. This might induce some performance reduction as *Tensor Cores* will not be available."
1997
                " For more details about this, or help on choosing the correct value for resizing, refer to this guide:"
1998
1999
2000
                " https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html#requirements-tc"
            )

2001
2002
2003
        if new_num_tokens is None:
            return old_embeddings

2004
2005
        is_quantized = hasattr(self, "hf_quantizer") and self.hf_quantizer is not None
        if is_deepspeed_zero3_enabled() and not is_quantized:
2006
2007
2008
2009
2010
2011
2012
            import deepspeed

            with deepspeed.zero.GatheredParameters(old_embeddings.weight, modifier_rank=None):
                old_num_tokens, old_embedding_dim = old_embeddings.weight.size()
        else:
            old_num_tokens, old_embedding_dim = old_embeddings.weight.size()

2013
        if old_num_tokens == new_num_tokens and not is_deepspeed_zero3_enabled():
2014
2015
            return old_embeddings

2016
2017
        if not isinstance(old_embeddings, nn.Embedding):
            raise TypeError(
Sylvain Gugger's avatar
Sylvain Gugger committed
2018
2019
2020
                f"Old embeddings are of type {type(old_embeddings)}, which is not an instance of {nn.Embedding}. You"
                " should either use a different resize function or make sure that `old_embeddings` are an instance of"
                f" {nn.Embedding}."
2021
2022
            )

2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
        # Build new embeddings

        # When using DeepSpeed ZeRO-3, we shouldn't create new embeddings with DeepSpeed init
        # because the shape of the new embedding layer is used across various modeling files
        # as well as to update config vocab size. Shape will be 0 when using DeepSpeed init leading
        # to errors when training.
        new_embeddings = nn.Embedding(
            new_num_tokens,
            old_embedding_dim,
            device=old_embeddings.weight.device,
            dtype=old_embeddings.weight.dtype,
        )

        # initialize all new embeddings (in particular added tokens)
        self._init_weights(new_embeddings)

        # Copy token embeddings from the previous weights

2041
2042
        # numbers of tokens to copy
        n = min(old_num_tokens, new_num_tokens)
2043

2044
        if is_deepspeed_zero3_enabled() and not is_quantized:
2045
2046
            import deepspeed

2047
2048
2049
            params = [old_embeddings.weight, new_embeddings.weight]
            with deepspeed.zero.GatheredParameters(params, modifier_rank=0):
                new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
2050
2051
        else:
            new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
2052
2053
2054

        return new_embeddings

2055
    def _get_resized_lm_head(
2056
2057
        self, old_lm_head: nn.Linear, new_num_tokens: Optional[int] = None, transposed: Optional[bool] = False
    ) -> nn.Linear:
2058
2059
2060
2061
2062
        """
        Build a resized Linear Module from a provided old Linear Module. Increasing the size will add newly initialized
        vectors at the end. Reducing the size will remove vectors from the end

        Args:
2063
            old_lm_head (`torch.nn.Linear`):
2064
                Old lm head liner layer to be resized.
2065
            new_num_tokens (`int`, *optional*):
2066
2067
2068
                New number of tokens in the linear matrix.

                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
2069
                vectors from the end. If not provided or `None`, just returns a pointer to the input tokens
2070
2071
2072
                `torch.nn.Linear` module of the model without doing anything. transposed (`bool`, *optional*, defaults
                to `False`): Whether `old_lm_head` is transposed or not. If True `old_lm_head.size()` is `lm_head_dim,
                vocab_size` else `vocab_size, lm_head_dim`.
2073
2074

        Return:
Sylvain Gugger's avatar
Sylvain Gugger committed
2075
2076
            `torch.nn.Linear`: Pointer to the resized Linear Module or the old Linear Module if `new_num_tokens` is
            `None`
2077
2078
2079
2080
        """
        if new_num_tokens is None:
            return old_lm_head

2081
2082
        is_quantized = hasattr(self, "hf_quantizer") and self.hf_quantizer is not None
        if is_deepspeed_zero3_enabled() and not is_quantized:
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
            import deepspeed

            with deepspeed.zero.GatheredParameters(old_lm_head.weight, modifier_rank=None):
                old_num_tokens, old_lm_head_dim = (
                    old_lm_head.weight.size() if not transposed else old_lm_head.weight.t().size()
                )
        else:
            old_num_tokens, old_lm_head_dim = (
                old_lm_head.weight.size() if not transposed else old_lm_head.weight.t().size()
            )
2093

2094
        if old_num_tokens == new_num_tokens and not is_deepspeed_zero3_enabled():
2095
2096
2097
2098
            return old_lm_head

        if not isinstance(old_lm_head, nn.Linear):
            raise TypeError(
Sylvain Gugger's avatar
Sylvain Gugger committed
2099
2100
2101
                f"Old language model head is of type {type(old_lm_head)}, which is not an instance of {nn.Linear}. You"
                " should either use a different resize function or make sure that `old_lm_head` are an instance of"
                f" {nn.Linear}."
2102
2103
2104
2105
2106
2107
            )

        # Build new lm head
        new_lm_head_shape = (old_lm_head_dim, new_num_tokens) if not transposed else (new_num_tokens, old_lm_head_dim)
        has_new_lm_head_bias = old_lm_head.bias is not None

2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
        # When using DeepSpeed ZeRO-3, we shouldn't create new embeddings with DeepSpeed init
        # because the shape of the new embedding layer is used across various modeling files
        # as well as to update config vocab size. Shape will be 0 when using DeepSpeed init leading
        # to errors when training.
        new_lm_head = nn.Linear(
            *new_lm_head_shape,
            bias=has_new_lm_head_bias,
            device=old_lm_head.weight.device,
            dtype=old_lm_head.weight.dtype,
        )

        # initialize new lm head (in particular added tokens)
        self._init_weights(new_lm_head)

2122
2123
        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)

2124
        if is_deepspeed_zero3_enabled() and not is_quantized:
2125
2126
            import deepspeed

2127
2128
            params = [old_lm_head.weight, old_lm_head.bias, new_lm_head.weight, new_lm_head.bias]
            with deepspeed.zero.GatheredParameters(params, modifier_rank=0):
2129
2130
2131
                self._copy_lm_head_original_to_resized(
                    new_lm_head, old_lm_head, num_tokens_to_copy, transposed, has_new_lm_head_bias
                )
2132
        else:
2133
2134
            self._copy_lm_head_original_to_resized(
                new_lm_head, old_lm_head, num_tokens_to_copy, transposed, has_new_lm_head_bias
2135
            )
2136
2137
2138

        return new_lm_head

2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
    def _copy_lm_head_original_to_resized(
        self, new_lm_head, old_lm_head, num_tokens_to_copy, transposed, has_new_lm_head_bias
    ):
        # Copy old lm head weights to new lm head
        if not transposed:
            new_lm_head.weight.data[:num_tokens_to_copy, :] = old_lm_head.weight.data[:num_tokens_to_copy, :]
        else:
            new_lm_head.weight.data[:, :num_tokens_to_copy] = old_lm_head.weight.data[:, :num_tokens_to_copy]

        # Copy bias weights to new lm head
        if has_new_lm_head_bias:
            new_lm_head.bias.data[:num_tokens_to_copy] = old_lm_head.bias.data[:num_tokens_to_copy]

2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
    def resize_position_embeddings(self, new_num_position_embeddings: int):
        raise NotImplementedError(
            f"`resize_position_embeddings` is not implemented for {self.__class__}`. To implement it, you should "
            f"overwrite this method in the class {self.__class__} in `modeling_{self.__class__.__module__}.py`"
        )

    def get_position_embeddings(self) -> Union[nn.Embedding, Tuple[nn.Embedding]]:
        raise NotImplementedError(
            f"`get_position_embeddings` is not implemented for {self.__class__}`. To implement it, you should "
            f"overwrite this method in the class {self.__class__} in `modeling_{self.__class__.__module__}.py`"
        )

2164
    def init_weights(self):
2165
        """
2166
2167
        If needed prunes and maybe initializes weights. If using a custom `PreTrainedModel`, you need to implement any
        initialization logic in `_init_weights`.
2168
        """
2169
2170
2171
2172
        # Prune heads if needed
        if self.config.pruned_heads:
            self.prune_heads(self.config.pruned_heads)

2173
2174
        if _init_weights:
            # Initialize weights
2175
            self.apply(self._initialize_weights)
2176
2177
2178
2179

            # Tie weights should be skipped when not initializing all weights
            # since from_pretrained(...) calls tie weights anyways
            self.tie_weights()
2180

2181
2182
2183
    def prune_heads(self, heads_to_prune: Dict[int, List[int]]):
        """
        Prunes heads of the base model.
2184

2185
        Arguments:
2186
            heads_to_prune (`Dict[int, List[int]]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2187
2188
2189
                Dictionary with keys being selected layer indices (`int`) and associated values being the list of heads
                to prune in said layer (list of `int`). For instance {1: [0, 2], 2: [2, 3]} will prune heads 0 and 2 on
                layer 1 and heads 2 and 3 on layer 2.
thomwolf's avatar
thomwolf committed
2190
        """
2191
        # save new sets of pruned heads as union of previously stored pruned heads and newly pruned heads
2192
        for layer, heads in heads_to_prune.items():
2193
2194
2195
            union_heads = set(self.config.pruned_heads.get(layer, [])) | set(heads)
            self.config.pruned_heads[layer] = list(union_heads)  # Unfortunately we have to store it as list for JSON

2196
        self.base_model._prune_heads(heads_to_prune)
thomwolf's avatar
thomwolf committed
2197

2198
    def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):
2199
2200
2201
2202
2203
        """
        Activates gradient checkpointing for the current model.

        Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint
        activations".
2204
2205
2206
2207
2208
2209
2210

        We pass the `__call__` method of the modules instead of `forward` because `__call__` attaches all the hooks of
        the module. https://discuss.pytorch.org/t/any-different-between-model-input-and-model-forward-input/3690/2

        Args:
            gradient_checkpointing_kwargs (dict, *optional*):
                Additional keyword arguments passed along to the `torch.utils.checkpoint.checkpoint` function.
2211
2212
2213
        """
        if not self.supports_gradient_checkpointing:
            raise ValueError(f"{self.__class__.__name__} does not support gradient checkpointing.")
2214
2215

        if gradient_checkpointing_kwargs is None:
2216
            gradient_checkpointing_kwargs = {"use_reentrant": True}
2217

2218
        gradient_checkpointing_func = functools.partial(checkpoint, **gradient_checkpointing_kwargs)
2219

2220
        # For old GC format (transformers < 4.35.0) for models that live on the Hub
Stas Bekman's avatar
Stas Bekman committed
2221
        # we will fall back to the overwritten `_set_gradient_checkpointing` method
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
        _is_using_old_format = "value" in inspect.signature(self._set_gradient_checkpointing).parameters

        if not _is_using_old_format:
            self._set_gradient_checkpointing(enable=True, gradient_checkpointing_func=gradient_checkpointing_func)
        else:
            self.apply(partial(self._set_gradient_checkpointing, value=True))
            logger.warn(
                "You are using an old version of the checkpointing format that is deprecated (We will also silently ignore `gradient_checkpointing_kwargs` in case you passed it)."
                "Please update to the new format on your modeling file. To use the new format, you need to completely remove the definition of the method `_set_gradient_checkpointing` in your model."
            )
2232

2233
2234
2235
2236
2237
2238
2239
        if getattr(self, "_hf_peft_config_loaded", False):
            # When using PEFT + gradient checkpointing + Trainer we need to make sure the input has requires_grad=True
            # we do it also on PEFT: https://github.com/huggingface/peft/blob/85013987aa82aa1af3da1236b6902556ce3e483e/src/peft/peft_model.py#L334
            # When training with PEFT, only LoRA layers will have requires grad set to True, but the output of frozen layers need to propagate
            # the gradients to make sure the gradient flows.
            self.enable_input_require_grads()

2240
    def _set_gradient_checkpointing(self, enable: bool = True, gradient_checkpointing_func: Callable = checkpoint):
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
        is_gradient_checkpointing_set = False

        # Apply it on the top-level module in case the top-level modules supports it
        # for example, LongT5Stack inherits from `PreTrainedModel`.
        if hasattr(self, "gradient_checkpointing"):
            self._gradient_checkpointing_func = gradient_checkpointing_func
            self.gradient_checkpointing = enable
            is_gradient_checkpointing_set = True

        for module in self.modules():
            if hasattr(module, "gradient_checkpointing"):
                module._gradient_checkpointing_func = gradient_checkpointing_func
                module.gradient_checkpointing = enable
                is_gradient_checkpointing_set = True

        if not is_gradient_checkpointing_set:
            raise ValueError(
                f"{self.__class__.__name__} is not compatible with gradient checkpointing. Make sure all the architecture support it by setting a boolean attribute"
                " `gradient_checkpointing` to modules of the model that uses checkpointing."
            )

2262
    def gradient_checkpointing_disable(self):
2263
2264
2265
2266
2267
2268
2269
        """
        Deactivates gradient checkpointing for the current model.

        Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint
        activations".
        """
        if self.supports_gradient_checkpointing:
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
            # For old GC format (transformers < 4.35.0) for models that live on the Hub
            # we will fall back to the overwritten `_set_gradient_checkpointing` methid
            _is_using_old_format = "value" in inspect.signature(self._set_gradient_checkpointing).parameters
            if not _is_using_old_format:
                self._set_gradient_checkpointing(enable=False)
            else:
                logger.warn(
                    "You are using an old version of the checkpointing format that is deprecated (We will also silently ignore `gradient_checkpointing_kwargs` in case you passed it)."
                    "Please update to the new format on your modeling file. To use the new format, you need to completely remove the definition of the method `_set_gradient_checkpointing` in your model."
                )
                self.apply(partial(self._set_gradient_checkpointing, value=False))
2281

2282
2283
2284
        if getattr(self, "_hf_peft_config_loaded", False):
            self.disable_input_require_grads()

2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
    @property
    def is_gradient_checkpointing(self) -> bool:
        """
        Whether gradient checkpointing is activated for this model or not.

        Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint
        activations".
        """
        return any(hasattr(m, "gradient_checkpointing") and m.gradient_checkpointing for m in self.modules())

2295
2296
2297
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
2298
        is_main_process: bool = True,
2299
2300
        state_dict: Optional[dict] = None,
        save_function: Callable = torch.save,
Sylvain Gugger's avatar
Sylvain Gugger committed
2301
        push_to_hub: bool = False,
2302
        max_shard_size: Union[int, str] = "5GB",
2303
        safe_serialization: bool = True,
2304
        variant: Optional[str] = None,
2305
        token: Optional[Union[str, bool]] = None,
2306
        save_peft_format: bool = True,
Sylvain Gugger's avatar
Sylvain Gugger committed
2307
        **kwargs,
2308
    ):
2309
2310
        """
        Save a model and its configuration file to a directory, so that it can be re-loaded using the
2311
        [`~PreTrainedModel.from_pretrained`] class method.
2312

2313
        Arguments:
2314
            save_directory (`str` or `os.PathLike`):
2315
                Directory to which to save. Will be created if it doesn't exist.
2316
2317
2318
2319
            is_main_process (`bool`, *optional*, defaults to `True`):
                Whether the process calling this is the main process or not. Useful when in distributed training like
                TPUs and 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.
2320
            state_dict (nested dictionary of `torch.Tensor`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2321
2322
2323
                The state dictionary of the model to save. Will default to `self.state_dict()`, but can be used to only
                save parts of the model or if special precautions need to be taken when recovering the state dictionary
                of a model (like when using model parallelism).
2324
            save_function (`Callable`):
2325
                The function to use to save the state dictionary. Useful on distributed training like TPUs when one
2326
2327
                need to replace `torch.save` by another method.
            push_to_hub (`bool`, *optional*, defaults to `False`):
2328
2329
2330
                Whether or not to push your model to the Hugging Face model 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).
2331
            max_shard_size (`int` or `str`, *optional*, defaults to `"5GB"`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2332
2333
                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 `"5MB"`).
2334
2335
                We default it to 5GB in order for models to be able to run easily on free-tier google colab instances
                without CPU OOM issues.
Sylvain Gugger's avatar
Sylvain Gugger committed
2336
2337
2338
2339
2340
2341
2342
2343

                <Tip warning={true}>

                If a single weight of the model is bigger than `max_shard_size`, it will be in its own checkpoint shard
                which will be bigger than `max_shard_size`.

                </Tip>

2344
            safe_serialization (`bool`, *optional*, defaults to `True`):
2345
                Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
2346
2347
            variant (`str`, *optional*):
                If specified, weights are saved in the format pytorch_model.<variant>.bin.
2348
2349
2350
            token (`str` or `bool`, *optional*):
                The token to use as HTTP bearer authorization for remote files. If `True`, or not specified, will use
                the token generated when running `huggingface-cli login` (stored in `~/.huggingface`).
2351
2352
2353
2354
            save_peft_format (`bool`, *optional*, defaults to `True`):
                For backward compatibility with PEFT library, in case adapter weights are attached to the model, all
                keys of the state dict of adapters needs to be pre-pended with `base_model.model`. Advanced users can
                disable this behaviours by setting `save_peft_format` to `False`.
2355
            kwargs (`Dict[str, Any]`, *optional*):
2356
                Additional key word arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
2357
        """
2358
        use_auth_token = kwargs.pop("use_auth_token", None)
2359
        ignore_metadata_errors = kwargs.pop("ignore_metadata_errors", False)
2360
2361
2362

        if use_auth_token is not None:
            warnings.warn(
2363
2364
                "The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
                FutureWarning,
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
            )
            if token is not None:
                raise ValueError(
                    "`token` and `use_auth_token` are both specified. Please set only the argument `token`."
                )
            token = use_auth_token

        if token is not None:
            kwargs["token"] = token

Younes Belkada's avatar
Younes Belkada committed
2375
2376
        _hf_peft_config_loaded = getattr(self, "_hf_peft_config_loaded", False)

2377
2378
2379
2380
        hf_quantizer = getattr(self, "hf_quantizer", None)
        quantization_serializable = (
            hf_quantizer is not None and isinstance(hf_quantizer, HfQuantizer) and hf_quantizer.is_serializable
        )
2381

2382
2383
2384
2385
        if hf_quantizer is not None and not _hf_peft_config_loaded and not quantization_serializable:
            raise ValueError(
                f"The model is quantized with {hf_quantizer.quantization_config.quant_method} and is not serializable - check out the warnings from"
                " the logger on the traceback to understand the reason why the quantized model is not serializable."
2386
2387
            )

2388
2389
2390
2391
2392
        if "save_config" in kwargs:
            warnings.warn(
                "`save_config` is deprecated and will be removed in v5 of Transformers. Use `is_main_process` instead."
            )
            is_main_process = kwargs.pop("save_config")
2393
2394
        if safe_serialization and not is_safetensors_available():
            raise ImportError("`safe_serialization` requires the `safetensors library: `pip install safetensors`.")
2395

2396
        if os.path.isfile(save_directory):
2397
            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
2398
            return
2399

2400
2401
        os.makedirs(save_directory, exist_ok=True)

2402
2403
        if push_to_hub:
            commit_message = kwargs.pop("commit_message", None)
2404
            repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
2405
            repo_id = self._create_repo(repo_id, **kwargs)
2406
            files_timestamps = self._get_files_timestamps(save_directory)
2407

Julien Chaumond's avatar
Julien Chaumond committed
2408
        # Only save the model itself if we are using distributed training
2409
        model_to_save = unwrap_model(self)
2410

2411
2412
2413
2414
2415
        # save the string version of dtype to the config, e.g. convert torch.float32 => "float32"
        # we currently don't use this setting automatically, but may start to use with v5
        dtype = get_parameter_dtype(model_to_save)
        model_to_save.config.torch_dtype = str(dtype).split(".")[1]

Julien Chaumond's avatar
Julien Chaumond committed
2416
2417
2418
        # Attach architecture to the config
        model_to_save.config.architectures = [model_to_save.__class__.__name__]

2419
2420
2421
2422
2423
        # If we have a custom model, we copy the file defining it in the folder and set the attributes so it can be
        # loaded from the Hub.
        if self._auto_class is not None:
            custom_object_save(self, save_directory, config=self.config)

2424
        # Save the config
2425
        if is_main_process:
2426
2427
            if not _hf_peft_config_loaded:
                model_to_save.config.save_pretrained(save_directory)
2428
            if self.can_generate():
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
                # generation config built from the model config + the model config holds generation kwargs -> generate
                # may revert to legacy behavior if the two don't match
                if (
                    model_to_save.generation_config._from_model_config
                    and model_to_save.config._has_non_default_generation_parameters()
                ):
                    new_generation_config = GenerationConfig.from_model_config(model_to_save.config)
                    if new_generation_config != model_to_save.generation_config:
                        logger.warning(
                            "Your generation config was originally created from the model config, but the model "
                            "config has changed since then. Unless you pass the `generation_config` argument to this "
                            "model's `generate` calls, they will revert to the legacy behavior where the base "
                            "`generate` parameterization is loaded from the model config instead. "
                            "To avoid this behavior and this warning, we recommend you to overwrite the generation "
                            "config model attribute before calling the model's `save_pretrained`, preferably also "
                            "removing any generation kwargs from the model config. This warning will be raised to an "
                            "exception in v4.41."
                        )
2447
                model_to_save.generation_config.save_pretrained(save_directory)
2448

2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
            if _hf_peft_config_loaded:
                logger.info(
                    "Detected adapters on the model, saving the model in the PEFT format, only adapter weights will be saved."
                )
                state_dict = model_to_save.get_adapter_state_dict()

                if save_peft_format:
                    logger.info(
                        "To match the expected format of the PEFT library, all keys of the state dict of adapters will be pre-pended with `base_model.model`."
                    )
                    peft_state_dict = {}
                    for key, value in state_dict.items():
                        peft_state_dict[f"base_model.model.{key}"] = value
                    state_dict = peft_state_dict

2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
                active_adapter = self.active_adapters()

                if len(active_adapter) > 1:
                    raise ValueError(
                        "Multiple active adapters detected, saving multiple active adapters is not supported yet. You can save adapters separately one by one "
                        "by iteratively calling `model.set_adapter(adapter_name)` then `model.save_pretrained(...)`"
                    )
                active_adapter = active_adapter[0]

                current_peft_config = self.peft_config[active_adapter]
2474
2475
                current_peft_config.save_pretrained(save_directory)

2476
2477
2478
        # Save the model
        if state_dict is None:
            state_dict = model_to_save.state_dict()
2479

2480
2481
2482
2483
2484
        # Translate state_dict from smp to hf if saving with smp >= 1.10
        if IS_SAGEMAKER_MP_POST_1_10:
            for smp_to_hf, _ in smp.state.module_manager.translate_functions:
                state_dict = smp_to_hf(state_dict)

2485
        # Handle the case where some state_dict keys shouldn't be saved
2486
        if self._keys_to_ignore_on_save is not None:
2487
            for ignore_key in self._keys_to_ignore_on_save:
2488
2489
                if ignore_key in state_dict.keys():
                    del state_dict[ignore_key]
2490
2491
2492
2493
2494
        if safe_serialization:
            # Safetensors does not allow tensor aliasing.
            # We're going to remove aliases before saving
            ptrs = collections.defaultdict(list)
            for name, tensor in state_dict.items():
2495
2496
2497
2498
2499
2500
2501
                # Sometimes in the state_dict we have non-tensor objects.
                # e.g. in bitsandbytes we have some `str` objects in the state_dict
                if isinstance(tensor, torch.Tensor):
                    ptrs[id_tensor_storage(tensor)].append(name)
                else:
                    # In the non-tensor case, fall back to the pointer of the object itself
                    ptrs[id(tensor)].append(name)
2502
2503
2504

            # These are all the pointers of shared tensors.
            shared_ptrs = {ptr: names for ptr, names in ptrs.items() if len(names) > 1}
2505
2506
2507
2508
            error_names = []
            to_delete_names = set()
            # Recursively descend to find tied weight keys
            _tied_weights_keys = _get_tied_weight_keys(self)
2509
2510
2511
            for names in shared_ptrs.values():
                # Removing the keys which are declared as known duplicates on
                # load. This allows to make sure the name which is kept is consistent.
2512
                if _tied_weights_keys is not None:
2513
2514
                    found = 0
                    for name in sorted(names):
2515
                        matches_pattern = any(re.search(pat, name) for pat in _tied_weights_keys)
2516
                        if matches_pattern and name in state_dict:
2517
2518
                            found += 1
                            if found < len(names):
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
                                to_delete_names.add(name)
            # We are entering a place where the weights and the transformers configuration do NOT match.
            shared_names, disjoint_names = _find_disjoint(shared_ptrs.values(), state_dict)
            # Those are actually tensor sharing but disjoint from each other, we can safely clone them
            # Reloaded won't have the same property, but it shouldn't matter in any meaningful way.
            for name in disjoint_names:
                state_dict[name] = state_dict[name].clone()

            # When not all duplicates have been cleaned, still remove those keys, but put a clear warning.
            # If the link between tensors was done at runtime then `from_pretrained` will not get
            # the key back leading to random tensor. A proper warning will be shown
            # during reload (if applicable), but since the file is not necessarily compatible with
            # the config, better show a proper warning.
            shared_names, identical_names = _find_identical(shared_names, state_dict)
            # delete tensors that have identical storage
            for inames in identical_names:
                known = inames.intersection(to_delete_names)
                for name in known:
                    del state_dict[name]
                unknown = inames.difference(to_delete_names)
                if len(unknown) > 1:
                    error_names.append(unknown)

            if shared_names:
                error_names.append(set(shared_names))

            if len(error_names) > 0:
                raise RuntimeError(
                    f"The weights trying to be saved contained shared tensors {error_names} that are mismatching the transformers base configuration. Try saving using `safe_serialization=False` or remove this tensor sharing.",
2548
                )
2549

Sylvain Gugger's avatar
Sylvain Gugger committed
2550
        # Shard the model if it is too big.
2551
2552
2553
2554
2555
        if not _hf_peft_config_loaded:
            weights_name = SAFE_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
            weights_name = _add_variant(weights_name, variant)
        else:
            weights_name = ADAPTER_SAFE_WEIGHTS_NAME if safe_serialization else ADAPTER_WEIGHTS_NAME
2556

2557
        shards, index = shard_checkpoint(state_dict, max_shard_size=max_shard_size, weights_name=weights_name)
Sylvain Gugger's avatar
Sylvain Gugger committed
2558
2559
2560
2561

        # Clean the folder from a previous save
        for filename in os.listdir(save_directory):
            full_filename = os.path.join(save_directory, filename)
2562
2563
            # If we have a shard file that is not going to be replaced, we delete it, but only from the main process
            # in distributed settings to avoid race conditions.
2564
            weights_no_suffix = weights_name.replace(".bin", "").replace(".safetensors", "")
2565
2566
2567

            # make sure that file to be deleted matches format of sharded file, e.g. pytorch_model-00001-of-00005
            filename_no_suffix = filename.replace(".bin", "").replace(".safetensors", "")
2568
            reg = re.compile(r"(.*?)-\d{5}-of-\d{5}")
2569

2570
            if (
2571
                filename.startswith(weights_no_suffix)
2572
2573
2574
                and os.path.isfile(full_filename)
                and filename not in shards.keys()
                and is_main_process
2575
                and reg.fullmatch(filename_no_suffix) is not None
2576
            ):
Sylvain Gugger's avatar
Sylvain Gugger committed
2577
                os.remove(full_filename)
2578

Sylvain Gugger's avatar
Sylvain Gugger committed
2579
2580
        # Save the model
        for shard_file, shard in shards.items():
2581
2582
2583
2584
2585
2586
            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, os.path.join(save_directory, shard_file), metadata={"format": "pt"})
            else:
                save_function(shard, os.path.join(save_directory, shard_file))
Sylvain Gugger's avatar
Sylvain Gugger committed
2587
2588

        if index is None:
2589
            path_to_weights = os.path.join(save_directory, weights_name)
2590
            logger.info(f"Model weights saved in {path_to_weights}")
Sylvain Gugger's avatar
Sylvain Gugger committed
2591
        else:
2592
            save_index_file = SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
2593
            save_index_file = os.path.join(save_directory, _add_variant(save_index_file, variant))
Sylvain Gugger's avatar
Sylvain Gugger committed
2594
2595
2596
2597
2598
2599
2600
2601
2602
            # 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(shards)} checkpoint shards. You can find where each parameters has been saved in the "
                f"index located at {save_index_file}."
            )
2603

Sylvain Gugger's avatar
Sylvain Gugger committed
2604
        if push_to_hub:
2605
2606
2607
2608
2609
2610
2611
2612
            # Eventually create an empty model card
            model_card = create_and_tag_model_card(
                repo_id, self.model_tags, token=token, ignore_metadata_errors=ignore_metadata_errors
            )

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

2613
            self._upload_modified_files(
2614
2615
2616
2617
                save_directory,
                repo_id,
                files_timestamps,
                commit_message=commit_message,
2618
                token=token,
2619
            )
Sylvain Gugger's avatar
Sylvain Gugger committed
2620

2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
    @wraps(PushToHubMixin.push_to_hub)
    def push_to_hub(self, *args, **kwargs):
        tags = self.model_tags if self.model_tags is not None else []

        tags_kwargs = kwargs.get("tags", [])
        if isinstance(tags_kwargs, str):
            tags_kwargs = [tags_kwargs]

        for tag in tags_kwargs:
            if tag not in tags:
                tags.append(tag)

        if tags:
            kwargs["tags"] = tags
        return super().push_to_hub(*args, **kwargs)

2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
    def get_memory_footprint(self, return_buffers=True):
        r"""
        Get the memory footprint of a model. This will return the memory footprint of the current model in bytes.
        Useful to benchmark the memory footprint of the current model and design some tests. Solution inspired from the
        PyTorch discussions: https://discuss.pytorch.org/t/gpu-memory-that-model-uses/56822/2

        Arguments:
            return_buffers (`bool`, *optional*, defaults to `True`):
                Whether to return the size of the buffer tensors in the computation of the memory footprint. Buffers
                are tensors that do not require gradients and not registered as parameters. E.g. mean and std in batch
                norm layers. Please see: https://discuss.pytorch.org/t/what-pytorch-means-by-buffers/120266/2
        """
        mem = sum([param.nelement() * param.element_size() for param in self.parameters()])
        if return_buffers:
            mem_bufs = sum([buf.nelement() * buf.element_size() for buf in self.buffers()])
            mem = mem + mem_bufs
        return mem

2655
    @wraps(torch.nn.Module.cuda)
2656
2657
    def cuda(self, *args, **kwargs):
        # Checks if the model has been loaded in 8-bit
Marc Sun's avatar
Marc Sun committed
2658
        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
2659
2660
2661
2662
2663
2664
2665
            raise ValueError(
                "Calling `cuda()` is not supported for `4-bit` or `8-bit` quantized models. Please use the model as it is, since the"
                " model has already been set to the correct devices and casted to the correct `dtype`."
            )
        else:
            return super().cuda(*args, **kwargs)

2666
    @wraps(torch.nn.Module.to)
2667
2668
    def to(self, *args, **kwargs):
        # Checks if the model has been loaded in 8-bit
Marc Sun's avatar
Marc Sun committed
2669
        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
2670
            raise ValueError(
Marc Sun's avatar
Marc Sun committed
2671
                "`.to` is not supported for `4-bit` or `8-bit` bitsandbytes models. Please use the model as it is, since the"
2672
2673
                " model has already been set to the correct devices and casted to the correct `dtype`."
            )
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
        elif getattr(self, "quantization_method", None) == QuantizationMethod.GPTQ:
            # For GPTQ models, we prevent users from casting the model to another dytpe to restrict unwanted behaviours.
            # the correct API should be to load the model with the desired dtype directly through `from_pretrained`.
            dtype_present_in_args = False

            if "dtype" not in kwargs:
                for arg in args:
                    if isinstance(arg, torch.dtype):
                        dtype_present_in_args = True
                        break
            else:
                dtype_present_in_args = True

            if dtype_present_in_args:
                raise ValueError(
                    "You cannot cast a GPTQ model in a new `dtype`. Make sure to load the model using `from_pretrained` using the desired"
                    " `dtype` by passing the correct `torch_dtype` argument."
                )
        return super().to(*args, **kwargs)
2693
2694

    def half(self, *args):
Marc Sun's avatar
Marc Sun committed
2695
        # Checks if the model is quantized
2696
        if getattr(self, "is_quantized", False):
2697
            raise ValueError(
Marc Sun's avatar
Marc Sun committed
2698
                "`.half()` is not supported for quantized model. Please use the model as it is, since the"
2699
2700
2701
2702
2703
2704
                " model has already been casted to the correct `dtype`."
            )
        else:
            return super().half(*args)

    def float(self, *args):
Marc Sun's avatar
Marc Sun committed
2705
        # Checks if the model is quantized
2706
        if getattr(self, "is_quantized", False):
2707
            raise ValueError(
Marc Sun's avatar
Marc Sun committed
2708
                "`.float()` is not supported for quantized model. Please use the model as it is, since the"
2709
2710
2711
2712
2713
                " model has already been casted to the correct `dtype`."
            )
        else:
            return super().float(*args)

2714
    @classmethod
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
    def from_pretrained(
        cls,
        pretrained_model_name_or_path: Optional[Union[str, os.PathLike]],
        *model_args,
        config: Optional[Union[PretrainedConfig, str, os.PathLike]] = None,
        cache_dir: Optional[Union[str, os.PathLike]] = None,
        ignore_mismatched_sizes: bool = False,
        force_download: bool = False,
        local_files_only: bool = False,
        token: Optional[Union[str, bool]] = None,
        revision: str = "main",
        use_safetensors: bool = None,
        **kwargs,
    ):
2729
2730
        r"""
        Instantiate a pretrained pytorch model from a pre-trained model configuration.
2731

Sylvain Gugger's avatar
Sylvain Gugger committed
2732
2733
        The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train
        the model, you should first set it back in training mode with `model.train()`.
2734

2735
        The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
2736
2737
        pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
        task.
2738

2739
        The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
2740
        weights are discarded.
2741

2742
        Parameters:
2743
            pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
2744
2745
                Can be either:

2746
2747
2748
2749
                    - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
                    - A path to a *directory* containing model weights saved using
                      [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`.
                    - A path or url to a *tensorflow index checkpoint file* (e.g, `./tf_model/model.ckpt.index`). In
Sylvain Gugger's avatar
Sylvain Gugger committed
2750
2751
2752
                      this case, `from_tf` should be set to `True` and a configuration object should be provided as
                      `config` argument. This loading path is slower than converting the TensorFlow checkpoint in a
                      PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
2753
                    - A path or url to a model folder containing a *flax checkpoint file* in *.msgpack* format (e.g,
Sylvain Gugger's avatar
Sylvain Gugger committed
2754
2755
                      `./flax_model/` containing `flax_model.msgpack`). In this case, `from_flax` should be set to
                      `True`.
2756
2757
2758
2759
2760
                    - `None` if you are both providing the configuration and state dictionary (resp. with keyword
                      arguments `config` and `state_dict`).
            model_args (sequence of positional arguments, *optional*):
                All remaining positional arguments will be passed to the underlying model's `__init__` method.
            config (`Union[PretrainedConfig, str, os.PathLike]`, *optional*):
2761
2762
                Can be either:

2763
2764
                    - an instance of a class derived from [`PretrainedConfig`],
                    - a string or path valid as input to [`~PretrainedConfig.from_pretrained`].
2765

2766
                Configuration for the model to use instead of an automatically loaded configuration. Configuration can
2767
2768
                be automatically loaded when:

2769
                    - The model is a model provided by the library (loaded with the *model id* string of a pretrained
2770
                      model).
Sylvain Gugger's avatar
Sylvain Gugger committed
2771
2772
                    - The model was saved using [`~PreTrainedModel.save_pretrained`] and is reloaded by supplying the
                      save directory.
2773
2774
2775
                    - The model is loaded by supplying a local directory as `pretrained_model_name_or_path` and a
                      configuration JSON file named *config.json* is found in the directory.
            state_dict (`Dict[str, torch.Tensor]`, *optional*):
2776
2777
2778
                A state dictionary to use instead of a state dictionary loaded from saved weights file.

                This option can be used if you want to create a model from a pretrained configuration but load your own
Sylvain Gugger's avatar
Sylvain Gugger committed
2779
                weights. In this case though, you should check if using [`~PreTrainedModel.save_pretrained`] and
2780
2781
                [`~PreTrainedModel.from_pretrained`] is not a simpler option.
            cache_dir (`Union[str, os.PathLike]`, *optional*):
2782
2783
                Path to a directory in which a downloaded pretrained model configuration should be cached if the
                standard cache should not be used.
2784
            from_tf (`bool`, *optional*, defaults to `False`):
2785
                Load the model weights from a TensorFlow checkpoint save file (see docstring of
2786
2787
                `pretrained_model_name_or_path` argument).
            from_flax (`bool`, *optional*, defaults to `False`):
2788
                Load the model weights from a Flax checkpoint save file (see docstring of
2789
2790
                `pretrained_model_name_or_path` argument).
            ignore_mismatched_sizes (`bool`, *optional*, defaults to `False`):
2791
2792
2793
                Whether or not to raise an error if some of the weights from the checkpoint do not have the same size
                as the weights of the model (if for instance, you are instantiating a model with 10 labels from a
                checkpoint with 3 labels).
2794
            force_download (`bool`, *optional*, defaults to `False`):
2795
2796
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
2797
            resume_download (`bool`, *optional*, defaults to `False`):
2798
2799
                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
                file exists.
2800
            proxies (`Dict[str, str]`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
2801
2802
                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
2803
            output_loading_info(`bool`, *optional*, defaults to `False`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2804
                Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
2805
            local_files_only(`bool`, *optional*, defaults to `False`):
Stas Bekman's avatar
Stas Bekman committed
2806
                Whether or not to only look at local files (i.e., do not try to download the model).
2807
            token (`str` or `bool`, *optional*):
2808
2809
                The token to use as HTTP bearer authorization for remote files. If `True`, or not specified, will use
                the token generated when running `huggingface-cli login` (stored in `~/.huggingface`).
2810
            revision (`str`, *optional*, defaults to `"main"`):
Julien Chaumond's avatar
Julien Chaumond committed
2811
                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
2812
                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
Julien Chaumond's avatar
Julien Chaumond committed
2813
                identifier allowed by git.
2814
2815
2816
2817
2818
2819
2820

                <Tip>

                To test a pull request you made on the Hub, you can pass `revision="refs/pr/<pr_number>".

                </Tip>

2821
            mirror (`str`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
2822
2823
2824
                Mirror source to accelerate downloads in China. If you are from China and have an accessibility
                problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
                Please refer to the mirror site for more information.
2825
            _fast_init(`bool`, *optional*, defaults to `True`):
2826
2827
                Whether or not to disable fast initialization.

2828
2829
                <Tip warning={true}>

Sylvain Gugger's avatar
Sylvain Gugger committed
2830
2831
2832
                One should only disable *_fast_init* to ensure backwards compatibility with `transformers.__version__ <
                4.6.0` for seeded model initialization. This argument will be removed at the next major version. See
                [pull request 11471](https://github.com/huggingface/transformers/pull/11471) for more information.
2833

2834
                </Tip>
2835
2836
            attn_implementation (`str`, *optional*):
                The attention implementation to use in the model (if relevant). Can be any of `"eager"` (manual implementation of the attention), `"sdpa"` (using [`F.scaled_dot_product_attention`](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention.html)), or `"flash_attention_2"` (using [Dao-AILab/flash-attention](https://github.com/Dao-AILab/flash-attention)). By default, if available, SDPA will be used for torch>=2.1.1. The default is otherwise the manual `"eager"` implementation.
2837

2838
2839
2840
            > Parameters for big model inference

            low_cpu_mem_usage(`bool`, *optional*):
2841
2842
2843
                Tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
                This is an experimental feature and a subject to change at any moment.
            torch_dtype (`str` or `torch.dtype`, *optional*):
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
                Override the default `torch.dtype` and load the model under a specific `dtype`. The different options
                are:

                1. `torch.float16` or `torch.bfloat16` or `torch.float`: load in a specified
                  `dtype`, ignoring the model's `config.torch_dtype` if one exists. If not specified
                  - the model will get loaded in `torch.float` (fp32).

                2. `"auto"` - A `torch_dtype` entry in the `config.json` file of the model will be
                  attempted to be used. If this entry isn't found then next check the `dtype` of the first weight in
                  the checkpoint that's of a floating point type and use that as `dtype`. This will load the model
                  using the `dtype` it was saved in at the end of the training. It can't be used as an indicator of how
                  the model was trained. Since it could be trained in one of half precision dtypes, but saved in fp32.

                <Tip>

                For some models the `dtype` they were trained in is unknown - you may try to check the model's paper or
                reach out to the authors and ask them to add this information to the model's card and to insert the
                `torch_dtype` entry in `config.json` on the hub.

                </Tip>

2865
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]` or `int` or `torch.device`, *optional*):
2866
2867
                A map that specifies where each submodule should go. It doesn't need to be refined to each
                parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the
2868
2869
2870
                same device. If we only pass the device (*e.g.*, `"cpu"`, `"cuda:1"`, `"mps"`, or a GPU ordinal rank
                like `1`) on which the model will be allocated, the device map will map the entire model to this
                device. Passing `device_map = 0` means put the whole model on GPU 0.
2871

2872
2873
                To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
                more information about each option see [designing a device
Patrick von Platen's avatar
Patrick von Platen committed
2874
                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
2875
2876
2877
            max_memory (`Dict`, *optional*):
                A dictionary device identifier to maximum memory. Will default to the maximum memory available for each
                GPU and the available CPU RAM if unset.
2878
2879
            offload_folder (`str` or `os.PathLike`, *optional*):
                If the `device_map` contains any value `"disk"`, the folder where we will offload weights.
2880
            offload_state_dict (`bool`, *optional*):
2881
                If `True`, will temporarily offload the CPU state dict to the hard drive to avoid getting out of CPU
2882
2883
                RAM if the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to
                `True` when there is some disk offload.
2884
2885
            offload_buffers (`bool`, *optional*):
                Whether or not to offload the buffers with the model parameters.
Marc Sun's avatar
Marc Sun committed
2886
2887
            quantization_config (`Union[QuantizationConfigMixin,Dict]`, *optional*):
                A dictionary of configuration parameters or a QuantizationConfigMixin object for quantization (e.g
2888
2889
2890
2891
                bitsandbytes, gptq). There may be other quantization-related kwargs, including `load_in_4bit` and
                `load_in_8bit`, which are parsed by QuantizationConfigParser. Supported only for bitsandbytes
                quantizations and not preferred. consider inserting all such arguments into quantization_config
                instead.
2892
2893
2894
            subfolder (`str`, *optional*, defaults to `""`):
                In case the relevant files are located inside a subfolder of the model repo on huggingface.co, you can
                specify the folder name here.
2895
2896
2897
            variant (`str`, *optional*):
                If specified load weights from `variant` filename, *e.g.* pytorch_model.<variant>.bin. `variant` is
                ignored when using `from_tf` or `from_flax`.
2898
2899
2900
            use_safetensors (`bool`, *optional*, defaults to `None`):
                Whether or not to use `safetensors` checkpoints. Defaults to `None`. If not specified and `safetensors`
                is not installed, it will be set to `False`.
2901

2902
            kwargs (remaining dictionary of keyword arguments, *optional*):
2903
                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
2904
                `output_attentions=True`). Behaves differently depending on whether a `config` is provided or
2905
2906
                automatically loaded:

2907
2908
                    - If a configuration is provided with `config`, `**kwargs` will be directly passed to the
                      underlying model's `__init__` method (we assume all relevant updates to the configuration have
2909
                      already been done)
2910
                    - If a configuration is not provided, `kwargs` will be first passed to the configuration class
Sylvain Gugger's avatar
Sylvain Gugger committed
2911
2912
2913
2914
                      initialization function ([`~PretrainedConfig.from_pretrained`]). Each key of `kwargs` that
                      corresponds to a configuration attribute will be used to override said attribute with the
                      supplied `kwargs` value. Remaining keys that do not correspond to any configuration attribute
                      will be passed to the underlying model's `__init__` function.
2915
2916
2917

        <Tip>

Sylvain Gugger's avatar
Sylvain Gugger committed
2918
2919
        Activate the special ["offline-mode"](https://huggingface.co/transformers/installation.html#offline-mode) to
        use this method in a firewalled environment.
2920
2921
2922
2923
2924
2925
2926

        </Tip>

        Examples:

        ```python
        >>> from transformers import BertConfig, BertModel
Sylvain Gugger's avatar
Sylvain Gugger committed
2927

2928
        >>> # Download model and configuration from huggingface.co and cache.
2929
        >>> model = BertModel.from_pretrained("google-bert/bert-base-uncased")
2930
        >>> # Model was saved using *save_pretrained('./test/saved_model/')* (for example purposes, not runnable).
Sylvain Gugger's avatar
Sylvain Gugger committed
2931
        >>> model = BertModel.from_pretrained("./test/saved_model/")
2932
        >>> # Update configuration during loading.
2933
        >>> model = BertModel.from_pretrained("google-bert/bert-base-uncased", output_attentions=True)
2934
2935
        >>> assert model.config.output_attentions == True
        >>> # Loading from a TF checkpoint file instead of a PyTorch model (slower, for example purposes, not runnable).
Sylvain Gugger's avatar
Sylvain Gugger committed
2936
2937
        >>> config = BertConfig.from_json_file("./tf_model/my_tf_model_config.json")
        >>> model = BertModel.from_pretrained("./tf_model/my_tf_checkpoint.ckpt.index", from_tf=True, config=config)
2938
        >>> # Loading from a Flax checkpoint file instead of a PyTorch model (slower)
2939
        >>> model = BertModel.from_pretrained("google-bert/bert-base-uncased", from_flax=True)
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
        ```

        * `low_cpu_mem_usage` algorithm:

        This is an experimental function that loads the model using ~1x model size CPU memory

        Here is how it works:

        1. save which state_dict keys we have
        2. drop state_dict before the model is created, since the latter takes 1x model size CPU memory
        3. after the model has been instantiated switch to the meta device all params/buffers that
        are going to be replaced from the loaded state_dict
        4. load state_dict 2nd time
        5. replace the params/buffers from the state_dict

        Currently, it can't handle deepspeed ZeRO stage 3 and ignores loading errors

        """
2958
2959
        state_dict = kwargs.pop("state_dict", None)
        from_tf = kwargs.pop("from_tf", False)
2960
        from_flax = kwargs.pop("from_flax", False)
2961
2962
2963
        resume_download = kwargs.pop("resume_download", False)
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
2964
        use_auth_token = kwargs.pop("use_auth_token", None)
2965
        trust_remote_code = kwargs.pop("trust_remote_code", None)
Sylvain Gugger's avatar
Sylvain Gugger committed
2966
        _ = kwargs.pop("mirror", None)
2967
2968
        from_pipeline = kwargs.pop("_from_pipeline", None)
        from_auto_class = kwargs.pop("_from_auto", False)
2969
        _fast_init = kwargs.pop("_fast_init", True)
2970
        torch_dtype = kwargs.pop("torch_dtype", None)
2971
2972
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", None)
        device_map = kwargs.pop("device_map", None)
2973
        max_memory = kwargs.pop("max_memory", None)
2974
        offload_folder = kwargs.pop("offload_folder", None)
2975
        offload_state_dict = kwargs.pop("offload_state_dict", False)
2976
        offload_buffers = kwargs.pop("offload_buffers", False)
2977
        load_in_8bit = kwargs.pop("load_in_8bit", False)
2978
        load_in_4bit = kwargs.pop("load_in_4bit", False)
2979
        quantization_config = kwargs.pop("quantization_config", None)
2980
        subfolder = kwargs.pop("subfolder", "")
2981
        commit_hash = kwargs.pop("_commit_hash", None)
2982
        variant = kwargs.pop("variant", None)
2983
        adapter_kwargs = kwargs.pop("adapter_kwargs", {})
2984
        adapter_name = kwargs.pop("adapter_name", "default")
2985
        use_flash_attention_2 = kwargs.pop("use_flash_attention_2", False)
2986

2987
2988
2989
        if is_fsdp_enabled():
            low_cpu_mem_usage = True

2990
2991
        if use_auth_token is not None:
            warnings.warn(
2992
2993
                "The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
                FutureWarning,
2994
2995
2996
2997
2998
2999
3000
            )
            if token is not None:
                raise ValueError(
                    "`token` and `use_auth_token` are both specified. Please set only the argument `token`."
                )
            token = use_auth_token

3001
        if token is not None and adapter_kwargs is not None and "token" not in adapter_kwargs:
3002
3003
            adapter_kwargs["token"] = token

3004
3005
        if use_safetensors is None and not is_safetensors_available():
            use_safetensors = False
3006
3007
3008
3009
3010
        if trust_remote_code is True:
            logger.warning(
                "The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is"
                " ignored."
            )
3011

3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
        if commit_hash is None:
            if not isinstance(config, PretrainedConfig):
                # We make a call to the config file first (which may be absent) to get the commit hash as soon as possible
                resolved_config_file = cached_file(
                    pretrained_model_name_or_path,
                    CONFIG_NAME,
                    cache_dir=cache_dir,
                    force_download=force_download,
                    resume_download=resume_download,
                    proxies=proxies,
                    local_files_only=local_files_only,
                    token=token,
                    revision=revision,
                    subfolder=subfolder,
3026
                    _raise_exceptions_for_gated_repo=False,
3027
3028
3029
3030
3031
3032
3033
                    _raise_exceptions_for_missing_entries=False,
                    _raise_exceptions_for_connection_errors=False,
                )
                commit_hash = extract_commit_hash(resolved_config_file, commit_hash)
            else:
                commit_hash = getattr(config, "_commit_hash", None)

3034
        if is_peft_available():
3035
3036
            _adapter_model_path = adapter_kwargs.pop("_adapter_model_path", None)

3037
3038
3039
3040
3041
3042
3043
3044
3045
            if _adapter_model_path is None:
                _adapter_model_path = find_adapter_config_file(
                    pretrained_model_name_or_path,
                    cache_dir=cache_dir,
                    force_download=force_download,
                    resume_download=resume_download,
                    proxies=proxies,
                    local_files_only=local_files_only,
                    _commit_hash=commit_hash,
3046
                    **adapter_kwargs,
3047
3048
                )
            if _adapter_model_path is not None and os.path.isfile(_adapter_model_path):
3049
                with open(_adapter_model_path, "r", encoding="utf-8") as f:
3050
                    _adapter_model_path = pretrained_model_name_or_path
3051
                    pretrained_model_name_or_path = json.load(f)["base_model_name_or_path"]
3052
3053
        else:
            _adapter_model_path = None
3054

3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
        # change device_map into a map if we passed an int, a str or a torch.device
        if isinstance(device_map, torch.device):
            device_map = {"": device_map}
        elif isinstance(device_map, str) and device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]:
            try:
                device_map = {"": torch.device(device_map)}
            except RuntimeError:
                raise ValueError(
                    "When passing device_map as a string, the value needs to be a device name (e.g. cpu, cuda:0) or "
                    f"'auto', 'balanced', 'balanced_low_0', 'sequential' but found {device_map}."
                )
        elif isinstance(device_map, int):
            if device_map < 0:
                raise ValueError(
                    "You can't pass device_map as a negative int. If you want to put the model on the cpu, pass device_map = 'cpu' "
                )
            else:
                device_map = {"": device_map}

3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
        if device_map is not None:
            if low_cpu_mem_usage is None:
                low_cpu_mem_usage = True
            elif not low_cpu_mem_usage:
                raise ValueError("Passing along a `device_map` requires `low_cpu_mem_usage=True`")

        if low_cpu_mem_usage:
            if is_deepspeed_zero3_enabled():
                raise ValueError(
                    "DeepSpeed Zero-3 is not compatible with `low_cpu_mem_usage=True` or with passing a `device_map`."
                )
            elif not is_accelerate_available():
                raise ImportError(
                    "Using `low_cpu_mem_usage=True` or a `device_map` requires Accelerate: `pip install accelerate`"
                )
3089

3090
3091
3092
        # handling bnb config from kwargs, remove after `load_in_{4/8}bit` deprecation.
        if load_in_4bit or load_in_8bit:
            if quantization_config is not None:
3093
                raise ValueError(
3094
                    "You can't pass `load_in_4bit`or `load_in_8bit` as a kwarg when passing "
3095
3096
3097
                    "`quantization_config` argument at the same time."
                )

3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
            # preparing BitsAndBytesConfig from kwargs
            config_dict = {k: v for k, v in kwargs.items() if k in inspect.signature(BitsAndBytesConfig).parameters}
            config_dict = {**config_dict, "load_in_4bit": load_in_4bit, "load_in_8bit": load_in_8bit}
            quantization_config, kwargs = BitsAndBytesConfig.from_dict(
                config_dict=config_dict, return_unused_kwargs=True, **kwargs
            )
            logger.warning(
                "The `load_in_4bit` and `load_in_8bit` arguments are deprecated and will be removed in the future versions. "
                "Please, pass a `BitsAndBytesConfig` object in `quantization_config` argument instead."
            )
3108

3109
        from_pt = not (from_tf | from_flax)
3110

3111
3112
3113
        user_agent = {"file_type": "model", "framework": "pytorch", "from_auto_class": from_auto_class}
        if from_pipeline is not None:
            user_agent["using_pipeline"] = from_pipeline
thomwolf's avatar
thomwolf committed
3114

3115
3116
3117
3118
        if is_offline_mode() and not local_files_only:
            logger.info("Offline mode: forcing local_files_only=True")
            local_files_only = True

3119
3120
3121
        # Load config if we don't provide a configuration
        if not isinstance(config, PretrainedConfig):
            config_path = config if config is not None else pretrained_model_name_or_path
3122
            config, model_kwargs = cls.config_class.from_pretrained(
3123
3124
3125
                config_path,
                cache_dir=cache_dir,
                return_unused_kwargs=True,
3126
                force_download=force_download,
3127
                resume_download=resume_download,
3128
                proxies=proxies,
3129
                local_files_only=local_files_only,
3130
                token=token,
Julien Chaumond's avatar
Julien Chaumond committed
3131
                revision=revision,
3132
                subfolder=subfolder,
3133
3134
                _from_auto=from_auto_class,
                _from_pipeline=from_pipeline,
3135
                **kwargs,
3136
3137
            )
        else:
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
            # In case one passes a config to `from_pretrained` + "attn_implementation"
            # override the `_attn_implementation` attribute to `attn_implementation` of the kwargs
            # Please see: https://github.com/huggingface/transformers/issues/28038

            # Overwrite `config._attn_implementation` by the one from the kwargs --> in auto-factory
            # we pop attn_implementation from the kwargs but this handles the case where users
            # passes manually the config to `from_pretrained`.
            config = copy.deepcopy(config)

            kwarg_attn_imp = kwargs.pop("attn_implementation", None)
            if kwarg_attn_imp is not None and config._attn_implementation != kwarg_attn_imp:
                config._attn_implementation = kwarg_attn_imp
3150
            model_kwargs = kwargs
3151

3152
3153
3154
3155
3156
        pre_quantized = getattr(config, "quantization_config", None) is not None
        if pre_quantized or quantization_config is not None:
            if pre_quantized:
                config.quantization_config = AutoHfQuantizer.merge_quantization_configs(
                    config.quantization_config, quantization_config
Marc Sun's avatar
Marc Sun committed
3157
3158
3159
                )
            else:
                config.quantization_config = quantization_config
3160
3161
3162
            hf_quantizer = AutoHfQuantizer.from_config(config.quantization_config, pre_quantized=pre_quantized)
        else:
            hf_quantizer = None
3163

3164
3165
3166
3167
3168
3169
        if hf_quantizer is not None:
            hf_quantizer.validate_environment(
                torch_dtype=torch_dtype, from_tf=from_tf, from_flax=from_flax, device_map=device_map
            )
            torch_dtype = hf_quantizer.update_torch_dtype(torch_dtype)
            device_map = hf_quantizer.update_device_map(device_map)
3170
3171
3172
3173

            # Force-set to `True` for more mem efficiency
            if low_cpu_mem_usage is None:
                low_cpu_mem_usage = True
3174
                logger.warning("`low_cpu_mem_usage` was None, now set to True since model is quantized.")
3175
        is_quantized = hf_quantizer is not None
3176

Sylvain Gugger's avatar
Sylvain Gugger committed
3177
3178
3179
3180
        # This variable will flag if we're loading a sharded checkpoint. In this case the archive file is just the
        # index of the files.
        is_sharded = False
        sharded_metadata = None
thomwolf's avatar
thomwolf committed
3181
        # Load model
Yih-Dar's avatar
Yih-Dar committed
3182
3183
        loading_info = None

3184
3185
3186
3187
        # Keep in fp32 modules
        keep_in_fp32_modules = None
        use_keep_in_fp32_modules = False

thomwolf's avatar
thomwolf committed
3188
        if pretrained_model_name_or_path is not None:
3189
            pretrained_model_name_or_path = str(pretrained_model_name_or_path)
3190
3191
            is_local = os.path.isdir(pretrained_model_name_or_path)
            if is_local:
3192
3193
3194
                if from_tf and os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, TF_WEIGHTS_NAME + ".index")
                ):
3195
                    # Load from a TF 1.0 checkpoint in priority if from_tf
3196
3197
3198
3199
                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, TF_WEIGHTS_NAME + ".index")
                elif from_tf and os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, TF2_WEIGHTS_NAME)
                ):
3200
                    # Load from a TF 2.0 checkpoint in priority if from_tf
3201
3202
3203
3204
                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, TF2_WEIGHTS_NAME)
                elif from_flax and os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)
                ):
3205
                    # Load from a Flax checkpoint in priority if from_flax
3206
                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)
3207
                elif use_safetensors is not False and os.path.isfile(
3208
                    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant))
3209
3210
                ):
                    # Load from a safetensors checkpoint
3211
3212
3213
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant)
                    )
3214
                elif use_safetensors is not False and os.path.isfile(
3215
3216
3217
                    os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
                    )
3218
3219
                ):
                    # Load from a sharded safetensors checkpoint
3220
3221
3222
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
                    )
3223
                    is_sharded = True
3224
3225
3226
                elif os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_NAME, variant))
                ):
thomwolf's avatar
thomwolf committed
3227
                    # Load from a PyTorch checkpoint
3228
3229
3230
3231
3232
3233
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_NAME, variant)
                    )
                elif os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_INDEX_NAME, variant))
                ):
Sylvain Gugger's avatar
Sylvain Gugger committed
3234
                    # Load from a sharded PyTorch checkpoint
3235
3236
3237
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_INDEX_NAME, variant)
                    )
Sylvain Gugger's avatar
Sylvain Gugger committed
3238
                    is_sharded = True
3239
3240
                # At this stage we don't have a weight file so we will raise an error.
                elif os.path.isfile(
3241
3242
                    os.path.join(pretrained_model_name_or_path, subfolder, TF_WEIGHTS_NAME + ".index")
                ) or os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, TF2_WEIGHTS_NAME)):
3243
                    raise EnvironmentError(
3244
3245
3246
                        f"Error no file named {_add_variant(WEIGHTS_NAME, variant)} found in directory"
                        f" {pretrained_model_name_or_path} but there is a file for TensorFlow weights. Use"
                        " `from_tf=True` to load this model from those weights."
3247
                    )
3248
                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)):
3249
                    raise EnvironmentError(
3250
3251
3252
                        f"Error no file named {_add_variant(WEIGHTS_NAME, variant)} found in directory"
                        f" {pretrained_model_name_or_path} but there is a file for Flax weights. Use `from_flax=True`"
                        " to load this model from those weights."
3253
                    )
3254
3255
3256
3257
3258
                elif use_safetensors:
                    raise EnvironmentError(
                        f"Error no file named {_add_variant(SAFE_WEIGHTS_NAME, variant)} found in directory"
                        f" {pretrained_model_name_or_path}."
                    )
thomwolf's avatar
thomwolf committed
3259
                else:
3260
                    raise EnvironmentError(
3261
3262
3263
                        f"Error no file named {_add_variant(WEIGHTS_NAME, variant)}, {TF2_WEIGHTS_NAME},"
                        f" {TF_WEIGHTS_NAME + '.index'} or {FLAX_WEIGHTS_NAME} found in directory"
                        f" {pretrained_model_name_or_path}."
3264
                    )
3265
            elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path)):
3266
                archive_file = pretrained_model_name_or_path
3267
                is_local = True
3268
            elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path + ".index")):
3269
3270
3271
3272
3273
                if not from_tf:
                    raise ValueError(
                        f"We found a TensorFlow checkpoint at {pretrained_model_name_or_path + '.index'}, please set "
                        "from_tf to True to load from this checkpoint."
                    )
3274
                archive_file = os.path.join(subfolder, pretrained_model_name_or_path + ".index")
3275
                is_local = True
3276
            elif is_remote_url(pretrained_model_name_or_path):
3277
                filename = pretrained_model_name_or_path
3278
                resolved_archive_file = download_url(pretrained_model_name_or_path)
3279
            else:
3280
3281
3282
3283
3284
                # set correct filename
                if from_tf:
                    filename = TF2_WEIGHTS_NAME
                elif from_flax:
                    filename = FLAX_WEIGHTS_NAME
3285
                elif use_safetensors is not False:
3286
                    filename = _add_variant(SAFE_WEIGHTS_NAME, variant)
3287
                else:
3288
                    filename = _add_variant(WEIGHTS_NAME, variant)
3289

3290
3291
                try:
                    # Load from URL or cache if already cached
3292
3293
3294
3295
3296
3297
                    cached_file_kwargs = {
                        "cache_dir": cache_dir,
                        "force_download": force_download,
                        "proxies": proxies,
                        "resume_download": resume_download,
                        "local_files_only": local_files_only,
3298
                        "token": token,
3299
3300
3301
                        "user_agent": user_agent,
                        "revision": revision,
                        "subfolder": subfolder,
3302
                        "_raise_exceptions_for_gated_repo": False,
3303
3304
3305
                        "_raise_exceptions_for_missing_entries": False,
                        "_commit_hash": commit_hash,
                    }
3306
                    resolved_archive_file = cached_file(pretrained_model_name_or_path, filename, **cached_file_kwargs)
3307

3308
                    # Since we set _raise_exceptions_for_missing_entries=False, we don't get an exception but a None
3309
                    # result when internet is up, the repo and revision exist, but the file does not.
3310
                    if resolved_archive_file is None and filename == _add_variant(SAFE_WEIGHTS_NAME, variant):
3311
3312
                        # Maybe the checkpoint is sharded, we try to grab the index name in this case.
                        resolved_archive_file = cached_file(
3313
3314
3315
                            pretrained_model_name_or_path,
                            _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant),
                            **cached_file_kwargs,
3316
3317
3318
                        )
                        if resolved_archive_file is not None:
                            is_sharded = True
3319
                        elif use_safetensors:
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
                            if revision == "main":
                                resolved_archive_file, revision, is_sharded = auto_conversion(
                                    pretrained_model_name_or_path, **cached_file_kwargs
                                )
                            cached_file_kwargs["revision"] = revision
                            if resolved_archive_file is None:
                                raise EnvironmentError(
                                    f"{pretrained_model_name_or_path} does not appear to have a file named"
                                    f" {_add_variant(SAFE_WEIGHTS_NAME, variant)} or {_add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)} "
                                    "and thus cannot be loaded with `safetensors`. Please make sure that the model has "
                                    "been saved with `safe_serialization=True` or do not set `use_safetensors=True`."
                                )
3332
3333
                        else:
                            # This repo has no safetensors file of any kind, we switch to PyTorch.
3334
                            filename = _add_variant(WEIGHTS_NAME, variant)
3335
                            resolved_archive_file = cached_file(
3336
                                pretrained_model_name_or_path, filename, **cached_file_kwargs
3337
                            )
3338
                    if resolved_archive_file is None and filename == _add_variant(WEIGHTS_NAME, variant):
Sylvain Gugger's avatar
Sylvain Gugger committed
3339
                        # Maybe the checkpoint is sharded, we try to grab the index name in this case.
3340
                        resolved_archive_file = cached_file(
3341
3342
3343
                            pretrained_model_name_or_path,
                            _add_variant(WEIGHTS_INDEX_NAME, variant),
                            **cached_file_kwargs,
3344
                        )
3345
3346
                        if resolved_archive_file is not None:
                            is_sharded = True
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379

                    if resolved_archive_file is not None:
                        if filename in [WEIGHTS_NAME, WEIGHTS_INDEX_NAME]:
                            # If the PyTorch file was found, check if there is a safetensors file on the repository
                            # If there is no safetensors file on the repositories, start an auto conversion
                            safe_weights_name = SAFE_WEIGHTS_INDEX_NAME if is_sharded else SAFE_WEIGHTS_NAME
                            has_file_kwargs = {
                                "revision": revision,
                                "proxies": proxies,
                                "token": token,
                            }
                            cached_file_kwargs = {
                                "cache_dir": cache_dir,
                                "force_download": force_download,
                                "resume_download": resume_download,
                                "local_files_only": local_files_only,
                                "user_agent": user_agent,
                                "subfolder": subfolder,
                                "_raise_exceptions_for_gated_repo": False,
                                "_raise_exceptions_for_missing_entries": False,
                                "_commit_hash": commit_hash,
                                **has_file_kwargs,
                            }
                            if not has_file(pretrained_model_name_or_path, safe_weights_name, **has_file_kwargs):
                                Thread(
                                    target=auto_conversion,
                                    args=(pretrained_model_name_or_path,),
                                    kwargs={"ignore_errors_during_conversion": True, **cached_file_kwargs},
                                    name="Thread-autoconversion",
                                ).start()
                    else:
                        # Otherwise, no PyTorch file was found, maybe there is a TF or Flax model file.
                        # We try those to give a helpful error message.
Sylvain Gugger's avatar
Sylvain Gugger committed
3380
3381
3382
                        has_file_kwargs = {
                            "revision": revision,
                            "proxies": proxies,
3383
                            "token": token,
Sylvain Gugger's avatar
Sylvain Gugger committed
3384
3385
3386
                        }
                        if has_file(pretrained_model_name_or_path, TF2_WEIGHTS_NAME, **has_file_kwargs):
                            raise EnvironmentError(
Sylvain Gugger's avatar
Sylvain Gugger committed
3387
                                f"{pretrained_model_name_or_path} does not appear to have a file named"
3388
3389
                                f" {_add_variant(WEIGHTS_NAME, variant)} but there is a file for TensorFlow weights."
                                " Use `from_tf=True` to load this model from those weights."
Sylvain Gugger's avatar
Sylvain Gugger committed
3390
3391
3392
                            )
                        elif has_file(pretrained_model_name_or_path, FLAX_WEIGHTS_NAME, **has_file_kwargs):
                            raise EnvironmentError(
Sylvain Gugger's avatar
Sylvain Gugger committed
3393
                                f"{pretrained_model_name_or_path} does not appear to have a file named"
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
                                f" {_add_variant(WEIGHTS_NAME, variant)} but there is a file for Flax weights. Use"
                                " `from_flax=True` to load this model from those weights."
                            )
                        elif variant is not None and has_file(
                            pretrained_model_name_or_path, WEIGHTS_NAME, **has_file_kwargs
                        ):
                            raise EnvironmentError(
                                f"{pretrained_model_name_or_path} does not appear to have a file named"
                                f" {_add_variant(WEIGHTS_NAME, variant)} but there is a file without the variant"
                                f" {variant}. Use `variant=None` to load this model from those weights."
Sylvain Gugger's avatar
Sylvain Gugger committed
3404
3405
3406
                            )
                        else:
                            raise EnvironmentError(
3407
3408
3409
                                f"{pretrained_model_name_or_path} does not appear to have a file named"
                                f" {_add_variant(WEIGHTS_NAME, variant)}, {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME} or"
                                f" {FLAX_WEIGHTS_NAME}."
Sylvain Gugger's avatar
Sylvain Gugger committed
3410
                            )
3411
3412
3413
3414
                except EnvironmentError:
                    # Raise any environment error raise by `cached_file`. It will have a helpful error message adapted
                    # to the original exception.
                    raise
3415
                except Exception as e:
3416
                    # For any other exception, we throw a generic error.
3417
                    raise EnvironmentError(
3418
3419
3420
                        f"Can't load the model for '{pretrained_model_name_or_path}'. If you were trying to load it"
                        " from 'https://huggingface.co/models', make sure you don't have a local directory with the"
                        f" same name. Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a"
3421
3422
                        f" directory containing a file named {_add_variant(WEIGHTS_NAME, variant)},"
                        f" {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME} or {FLAX_WEIGHTS_NAME}."
3423
                    ) from e
3424

3425
            if is_local:
3426
                logger.info(f"loading weights file {archive_file}")
3427
                resolved_archive_file = archive_file
3428
            else:
3429
                logger.info(f"loading weights file {filename} from cache at {resolved_archive_file}")
3430
        else:
thomwolf's avatar
thomwolf committed
3431
            resolved_archive_file = None
3432

Sylvain Gugger's avatar
Sylvain Gugger committed
3433
3434
        # We'll need to download and cache each checkpoint shard if the checkpoint is sharded.
        if is_sharded:
3435
            # rsolved_archive_file becomes a list of files that point to the different checkpoint shards in this case.
Sylvain Gugger's avatar
Sylvain Gugger committed
3436
3437
3438
3439
3440
3441
3442
3443
            resolved_archive_file, sharded_metadata = get_checkpoint_shard_files(
                pretrained_model_name_or_path,
                resolved_archive_file,
                cache_dir=cache_dir,
                force_download=force_download,
                proxies=proxies,
                resume_download=resume_download,
                local_files_only=local_files_only,
3444
                token=token,
Sylvain Gugger's avatar
Sylvain Gugger committed
3445
3446
                user_agent=user_agent,
                revision=revision,
3447
                subfolder=subfolder,
3448
                _commit_hash=commit_hash,
Sylvain Gugger's avatar
Sylvain Gugger committed
3449
3450
            )

3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
        if (
            is_safetensors_available()
            and isinstance(resolved_archive_file, str)
            and resolved_archive_file.endswith(".safetensors")
        ):
            with safe_open(resolved_archive_file, framework="pt") as f:
                metadata = f.metadata()

            if metadata.get("format") == "pt":
                pass
            elif metadata.get("format") == "tf":
                from_tf = True
                logger.info("A TensorFlow safetensors file is being loaded in a PyTorch model.")
            elif metadata.get("format") == "flax":
                from_flax = True
                logger.info("A Flax safetensors file is being loaded in a PyTorch model.")
3467
3468
3469
            elif metadata.get("format") == "mlx":
                # This is a mlx file, we assume weights are compatible with pt
                pass
3470
3471
            else:
                raise ValueError(
3472
                    f"Incompatible safetensors file. File metadata is not ['pt', 'tf', 'flax', 'mlx'] but {metadata.get('format')}"
3473
3474
3475
3476
                )

        from_pt = not (from_tf | from_flax)

3477
3478
        # load pt weights early so that we know which dtype to init the model under
        if from_pt:
3479
            if not is_sharded and state_dict is None:
Sylvain Gugger's avatar
Sylvain Gugger committed
3480
3481
                # Time to load the checkpoint
                state_dict = load_state_dict(resolved_archive_file)
3482

3483
3484
3485
            # set dtype to instantiate the model under:
            # 1. If torch_dtype is not None, we use that dtype
            # 2. If torch_dtype is "auto", we auto-detect dtype from the loaded state_dict, by checking its first
3486
            #    weights entry that is of a floating type - we assume all floating dtype weights are of the same dtype
3487
3488
            # we also may have config.torch_dtype available, but we won't rely on it till v5
            dtype_orig = None
3489

3490
3491
3492
            if torch_dtype is not None:
                if isinstance(torch_dtype, str):
                    if torch_dtype == "auto":
3493
3494
3495
                        if hasattr(config, "torch_dtype") and config.torch_dtype is not None:
                            torch_dtype = config.torch_dtype
                            logger.info(f"Will use torch_dtype={torch_dtype} as defined in model's config object")
Sylvain Gugger's avatar
Sylvain Gugger committed
3496
                        else:
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
                            if is_sharded and "dtype" in sharded_metadata:
                                torch_dtype = sharded_metadata["dtype"]
                            elif not is_sharded:
                                torch_dtype = get_state_dict_dtype(state_dict)
                            else:
                                one_state_dict = load_state_dict(resolved_archive_file[0])
                                torch_dtype = get_state_dict_dtype(one_state_dict)
                                del one_state_dict  # free CPU memory
                            logger.info(
                                "Since the `torch_dtype` attribute can't be found in model's config object, "
                                "will use torch_dtype={torch_dtype} as derived from model's weights"
                            )
3509
3510
                    else:
                        raise ValueError(
3511
                            f'`torch_dtype` can be either `torch.dtype` or `"auto"`, but received {torch_dtype}'
3512
3513
3514
                        )
                dtype_orig = cls._set_default_torch_dtype(torch_dtype)

3515
            # Check if `_keep_in_fp32_modules` is not None
3516
            use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and (
3517
                (torch_dtype == torch.float16) or hasattr(hf_quantizer, "use_keep_in_fp32_modules")
3518
3519
            )

3520
3521
3522
            if is_sharded:
                loaded_state_dict_keys = sharded_metadata["all_checkpoint_keys"]
            else:
3523
                loaded_state_dict_keys = list(state_dict.keys())
3524
3525
3526
3527
            if low_cpu_mem_usage or (use_keep_in_fp32_modules and is_accelerate_available()):
                # In case some weights need to be kept in float32 and accelerate is not installed,
                # we later on want to take the path where state_dict is not None, that is the one
                # that do not require accelerate.
3528
                state_dict = None
3529

3530
3531
        config.name_or_path = pretrained_model_name_or_path

3532
        # Instantiate model.
3533
3534
        init_contexts = [no_init_weights(_enable=_fast_init)]

3535
        if is_deepspeed_zero3_enabled() and not is_quantized:
3536
3537
3538
            import deepspeed

            logger.info("Detected DeepSpeed ZeRO-3: activating zero.init() for this model")
3539
            init_contexts = [deepspeed.zero.Init(config_dict_or_path=deepspeed_config())] + init_contexts
3540
        elif low_cpu_mem_usage:
3541
3542
            init_contexts.append(init_empty_weights())

3543
3544
3545
3546
        config = copy.deepcopy(config)  # We do not want to modify the config inplace in from_pretrained.
        config = cls._autoset_attn_implementation(
            config, use_flash_attention_2=use_flash_attention_2, torch_dtype=torch_dtype, device_map=device_map
        )
3547

3548
        with ContextManagers(init_contexts):
3549
            # Let's make sure we don't run the init function of buffer modules
3550
3551
            model = cls(config, *model_args, **model_kwargs)

3552
3553
3554
        # make sure we use the model's config since the __init__ call might have copied it
        config = model.config

3555
3556
        # Check first if we are `from_pt`
        if use_keep_in_fp32_modules:
3557
            if is_accelerate_available() and not is_deepspeed_zero3_enabled():
3558
                low_cpu_mem_usage = True
3559
3560
3561
3562
            keep_in_fp32_modules = model._keep_in_fp32_modules
        else:
            keep_in_fp32_modules = []

3563
3564
3565
        if hf_quantizer is not None:
            hf_quantizer.preprocess_model(
                model=model, device_map=device_map, keep_in_fp32_modules=keep_in_fp32_modules
3566
            )
3567

3568
3569
3570
3571
3572
3573
            # We store the original dtype for quantized models as we cannot easily retrieve it
            # once the weights have been quantized
            # Note that once you have loaded a quantized model, you can't change its dtype so this will
            # remain a single source of truth
            config._pre_quantization_dtype = torch_dtype

3574
        if isinstance(device_map, str):
3575
            special_dtypes = {}
3576
3577
3578

            if hf_quantizer is not None:
                special_dtypes.update(hf_quantizer.get_special_dtypes_update(model, torch_dtype))
3579
3580
3581
3582
3583
3584
3585
3586
3587

            special_dtypes.update(
                {
                    name: torch.float32
                    for name, _ in model.named_parameters()
                    if any(m in name for m in keep_in_fp32_modules)
                }
            )

3588
3589
            target_dtype = torch_dtype

3590
3591
            if hf_quantizer is not None:
                target_dtype = hf_quantizer.adjust_target_dtype(target_dtype)
3592

Marc Sun's avatar
Marc Sun committed
3593
            no_split_modules = model._get_no_split_modules(device_map)
3594
3595
3596
3597
3598
            if device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]:
                raise ValueError(
                    "If passing a string for `device_map`, please choose 'auto', 'balanced', 'balanced_low_0' or "
                    "'sequential'."
                )
3599

3600
            device_map_kwargs = {"no_split_module_classes": no_split_modules}
3601
            if "special_dtypes" in inspect.signature(infer_auto_device_map).parameters:
3602
                device_map_kwargs["special_dtypes"] = special_dtypes
3603
            elif len(special_dtypes) > 0:
3604
                logger.warning(
3605
3606
3607
                    "This model has some weights that should be kept in higher precision, you need to upgrade "
                    "`accelerate` to properly deal with them (`pip install --upgrade accelerate`)."
                )
3608
            if device_map != "sequential":
3609
3610
                max_memory = get_balanced_memory(
                    model,
3611
                    dtype=target_dtype,
3612
                    low_zero=(device_map == "balanced_low_0"),
3613
                    max_memory=max_memory,
3614
                    **device_map_kwargs,
3615
                )
Marc Sun's avatar
Marc Sun committed
3616
3617
            else:
                max_memory = get_max_memory(max_memory)
3618
3619
            if hf_quantizer is not None:
                max_memory = hf_quantizer.adjust_max_memory(max_memory)
3620
            device_map_kwargs["max_memory"] = max_memory
Marc Sun's avatar
Marc Sun committed
3621

3622
3623
            # Make sure tied weights are tied before creating the device map.
            model.tie_weights()
3624
            device_map = infer_auto_device_map(model, dtype=target_dtype, **device_map_kwargs)
3625

3626
3627
            if hf_quantizer is not None:
                hf_quantizer.validate_environment(device_map=device_map)
3628

3629
3630
3631
3632
        elif device_map is not None:
            model.tie_weights()
            tied_params = find_tied_parameters(model)
            # check if we don't have tied param in different devices
3633
            check_tied_parameters_on_same_device(tied_params, device_map)
3634

3635
        if from_tf:
3636
            if resolved_archive_file.endswith(".index"):
3637
3638
3639
3640
3641
                # Load from a TensorFlow 1.X checkpoint - provided by original authors
                model = cls.load_tf_weights(model, config, resolved_archive_file[:-6])  # Remove the '.index'
            else:
                # Load from our TensorFlow 2.0 checkpoints
                try:
3642
                    from .modeling_tf_pytorch_utils import load_tf2_checkpoint_in_pytorch_model
3643

Yih-Dar's avatar
Yih-Dar committed
3644
3645
3646
                    model, loading_info = load_tf2_checkpoint_in_pytorch_model(
                        model, resolved_archive_file, allow_missing_keys=True, output_loading_info=True
                    )
3647
                except ImportError:
3648
                    logger.error(
Sylvain Gugger's avatar
Sylvain Gugger committed
3649
3650
3651
                        "Loading a TensorFlow model in PyTorch, requires both PyTorch and TensorFlow to be installed."
                        " Please see https://pytorch.org/ and https://www.tensorflow.org/install/ for installation"
                        " instructions."
3652
                    )
3653
                    raise
3654
3655
3656
3657
3658
3659
3660
        elif from_flax:
            try:
                from .modeling_flax_pytorch_utils import load_flax_checkpoint_in_pytorch_model

                model = load_flax_checkpoint_in_pytorch_model(model, resolved_archive_file)
            except ImportError:
                logger.error(
Sylvain Gugger's avatar
Sylvain Gugger committed
3661
3662
3663
                    "Loading a Flax model in PyTorch, requires both PyTorch and Flax to be installed. Please see"
                    " https://pytorch.org/ and https://flax.readthedocs.io/en/latest/installation.html for"
                    " installation instructions."
3664
3665
                )
                raise
3666
        elif from_pt:
3667
3668
3669
            # restore default dtype
            if dtype_orig is not None:
                torch.set_default_dtype(dtype_orig)
Sylvain Gugger's avatar
Sylvain Gugger committed
3670
3671
3672
3673
3674
3675
3676
3677
            (
                model,
                missing_keys,
                unexpected_keys,
                mismatched_keys,
                offload_index,
                error_msgs,
            ) = cls._load_pretrained_model(
3678
3679
3680
3681
3682
3683
3684
3685
3686
                model,
                state_dict,
                loaded_state_dict_keys,  # XXX: rename?
                resolved_archive_file,
                pretrained_model_name_or_path,
                ignore_mismatched_sizes=ignore_mismatched_sizes,
                sharded_metadata=sharded_metadata,
                _fast_init=_fast_init,
                low_cpu_mem_usage=low_cpu_mem_usage,
3687
3688
3689
3690
                device_map=device_map,
                offload_folder=offload_folder,
                offload_state_dict=offload_state_dict,
                dtype=torch_dtype,
3691
                hf_quantizer=hf_quantizer,
3692
                keep_in_fp32_modules=keep_in_fp32_modules,
3693
            )
3694

3695
3696
        # make sure token embedding weights are still tied if needed
        model.tie_weights()
3697

3698
        # Set model in evaluation mode to deactivate DropOut modules by default
3699
3700
        model.eval()

3701
        # If it is a model with generation capabilities, attempt to load the generation config
3702
        if model.can_generate() and pretrained_model_name_or_path is not None:
3703
3704
3705
3706
3707
3708
3709
3710
            try:
                model.generation_config = GenerationConfig.from_pretrained(
                    pretrained_model_name_or_path,
                    cache_dir=cache_dir,
                    force_download=force_download,
                    resume_download=resume_download,
                    proxies=proxies,
                    local_files_only=local_files_only,
3711
                    token=token,
3712
3713
3714
3715
3716
3717
                    revision=revision,
                    subfolder=subfolder,
                    _from_auto=from_auto_class,
                    _from_pipeline=from_pipeline,
                    **kwargs,
                )
3718
            except OSError:
3719
3720
3721
3722
3723
                logger.info(
                    "Generation config file not found, using a generation config created from the model config."
                )
                pass

3724
3725
        # Dispatch model with hooks on all devices if necessary
        if device_map is not None:
3726
3727
3728
3729
            device_map_kwargs = {
                "device_map": device_map,
                "offload_dir": offload_folder,
                "offload_index": offload_index,
3730
                "offload_buffers": offload_buffers,
3731
            }
3732
            if "skip_keys" in inspect.signature(dispatch_model).parameters:
3733
                device_map_kwargs["skip_keys"] = model._skip_keys_device_placement
3734
3735
            if not is_fsdp_enabled() and not is_deepspeed_zero3_enabled():
                dispatch_model(model, **device_map_kwargs)
3736

3737
3738
3739
        if hf_quantizer is not None:
            hf_quantizer.postprocess_model(model)
            model.hf_quantizer = hf_quantizer
Marc Sun's avatar
Marc Sun committed
3740

3741
        if _adapter_model_path is not None:
3742
            model.load_adapter(
3743
                _adapter_model_path,
3744
3745
                adapter_name=adapter_name,
                token=token,
3746
                adapter_kwargs=adapter_kwargs,
3747
3748
            )

thomwolf's avatar
thomwolf committed
3749
        if output_loading_info:
Yih-Dar's avatar
Yih-Dar committed
3750
3751
3752
3753
3754
3755
3756
            if loading_info is None:
                loading_info = {
                    "missing_keys": missing_keys,
                    "unexpected_keys": unexpected_keys,
                    "mismatched_keys": mismatched_keys,
                    "error_msgs": error_msgs,
                }
thomwolf's avatar
thomwolf committed
3757
3758
            return model, loading_info

3759
3760
        return model

3761
    @classmethod
Sylvain Gugger's avatar
Sylvain Gugger committed
3762
3763
3764
3765
    def _load_pretrained_model(
        cls,
        model,
        state_dict,
3766
        loaded_keys,
Sylvain Gugger's avatar
Sylvain Gugger committed
3767
3768
3769
3770
3771
        resolved_archive_file,
        pretrained_model_name_or_path,
        ignore_mismatched_sizes=False,
        sharded_metadata=None,
        _fast_init=True,
3772
        low_cpu_mem_usage=False,
3773
3774
        device_map=None,
        offload_folder=None,
3775
        offload_state_dict=None,
3776
        dtype=None,
3777
        hf_quantizer=None,
3778
        keep_in_fp32_modules=None,
3779
    ):
Sylvain Gugger's avatar
Sylvain Gugger committed
3780
        is_safetensors = False
3781
        is_quantized = hf_quantizer is not None
3782

Sylvain Gugger's avatar
Sylvain Gugger committed
3783
        if device_map is not None and "disk" in device_map.values():
Sylvain Gugger's avatar
Sylvain Gugger committed
3784
3785
3786
3787
3788
            archive_file = (
                resolved_archive_file[0] if isinstance(resolved_archive_file, (list, tuple)) else resolved_archive_file
            )
            is_safetensors = archive_file.endswith(".safetensors")
            if offload_folder is None and not is_safetensors:
Sylvain Gugger's avatar
Sylvain Gugger committed
3789
3790
                raise ValueError(
                    "The current `device_map` had weights offloaded to the disk. Please provide an `offload_folder`"
Sylvain Gugger's avatar
Sylvain Gugger committed
3791
3792
                    " for them. Alternatively, make sure you have `safetensors` installed if the model you are using"
                    " offers the weights in this format."
Sylvain Gugger's avatar
Sylvain Gugger committed
3793
                )
Sylvain Gugger's avatar
Sylvain Gugger committed
3794
3795
            if offload_folder is not None:
                os.makedirs(offload_folder, exist_ok=True)
3796
3797
3798
            if offload_state_dict is None:
                offload_state_dict = True

3799
        is_sharded_safetensors = is_safetensors and sharded_metadata is not None
Patrick von Platen's avatar
Patrick von Platen committed
3800
3801
3802
3803

        # tie the model weights before retrieving the state_dict
        model.tie_weights()

3804
        # Retrieve missing & unexpected_keys
3805
3806
        model_state_dict = model.state_dict()
        expected_keys = list(model_state_dict.keys())
3807
3808
        prefix = model.base_model_prefix

Sylvain Gugger's avatar
Sylvain Gugger committed
3809
3810
3811
3812
3813
3814
3815
        def _fix_key(key):
            if "beta" in key:
                return key.replace("beta", "bias")
            if "gamma" in key:
                return key.replace("gamma", "weight")
            return key

3816
        original_loaded_keys = loaded_keys
Sylvain Gugger's avatar
Sylvain Gugger committed
3817
3818
        loaded_keys = [_fix_key(key) for key in loaded_keys]

3819
3820
3821
3822
3823
3824
        if len(prefix) > 0:
            has_prefix_module = any(s.startswith(prefix) for s in loaded_keys)
            expects_prefix_module = any(s.startswith(prefix) for s in expected_keys)
        else:
            has_prefix_module = False
            expects_prefix_module = False
Patrick von Platen's avatar
Patrick von Platen committed
3825
3826
3827

        # key re-naming operations are never done on the keys
        # that are loaded, but always on the keys of the newly initialized model
3828
3829
        remove_prefix_from_model = not has_prefix_module and expects_prefix_module
        add_prefix_to_model = has_prefix_module and not expects_prefix_module
3830

3831
        if remove_prefix_from_model:
3832
3833
3834
            _prefix = f"{prefix}."
            expected_keys_not_prefixed = [s for s in expected_keys if not s.startswith(_prefix)]
            expected_keys = [s[len(_prefix) :] if s.startswith(_prefix) else s for s in expected_keys]
3835
        elif add_prefix_to_model:
3836
3837
            expected_keys = [".".join([prefix, s]) for s in expected_keys]

3838
        missing_keys = sorted(set(expected_keys) - set(loaded_keys))
Sylvain Gugger's avatar
Sylvain Gugger committed
3839
3840
3841
3842
3843
3844
3845
3846
        unexpected_keys = set(loaded_keys) - set(expected_keys)
        # Remove nonpersistent buffers from unexpected keys: they are not in the state dict but will be in the model
        # buffers
        model_buffers = {n for n, _ in model.named_buffers()}
        if remove_prefix_from_model:
            model_buffers = {key[len(_prefix) :] if key.startswith(_prefix) else key for key in model_buffers}
        elif add_prefix_to_model:
            model_buffers = {".".join([prefix, key]) for key in model_buffers}
3847
        unexpected_keys = sorted(unexpected_keys - model_buffers)
3848

3849
        model.tie_weights()
3850
        if device_map is None and not is_fsdp_enabled() and not is_deepspeed_zero3_enabled():
3851
3852
3853
3854
            ptrs = collections.defaultdict(list)
            for name, tensor in model.state_dict().items():
                id_tensor = id_tensor_storage(tensor)
                ptrs[id_tensor].append(name)
Sylvain Gugger's avatar
Sylvain Gugger committed
3855

3856
3857
3858
3859
3860
            # These are all the pointers of shared tensors.
            tied_params = [names for _, names in ptrs.items() if len(names) > 1]
        else:
            # id function doesn't work for meta tensor so we need this function
            tied_params = find_tied_parameters(model)
Sylvain Gugger's avatar
Sylvain Gugger committed
3861
3862

        for group in tied_params:
Sylvain Gugger's avatar
Sylvain Gugger committed
3863
3864
3865
3866
            if remove_prefix_from_model:
                group = [key[len(_prefix) :] if key.startswith(_prefix) else key for key in group]
            elif add_prefix_to_model:
                group = [".".join([prefix, key]) for key in group]
Sylvain Gugger's avatar
Sylvain Gugger committed
3867
3868
3869
            missing_in_group = [k for k in missing_keys if k in group]
            if len(missing_in_group) > 0 and len(missing_in_group) < len(group):
                missing_keys = [k for k in missing_keys if k not in missing_in_group]
3870

3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
        # 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_missing is not None:
            for pat in cls._keys_to_ignore_on_load_missing:
                missing_keys = [k for k in missing_keys if re.search(pat, k) is None]

        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]

3881
3882
3883
        if hf_quantizer is not None:
            missing_keys = hf_quantizer.update_missing_keys(model, missing_keys, prefix)

3884
3885
3886
3887
        # retrieve weights on meta device and put them back on CPU.
        # This is not ideal in terms of memory, but if we don't do that not, we can't initialize them in the next step
        if low_cpu_mem_usage:
            for key in missing_keys:
Susnato Dhar's avatar
Susnato Dhar committed
3888
3889
                if key in list(model_state_dict.keys()):
                    key = key
3890
3891
                elif f"{prefix}.{key}" in list(model_state_dict.keys()):
                    key = f"{prefix}.{key}"
Susnato Dhar's avatar
Susnato Dhar committed
3892
                elif key.startswith(prefix) and ".".join(key.split(".")[1:]) in list(model_state_dict.keys()):
3893
3894
                    key = ".".join(key.split(".")[1:])
                param = model_state_dict[key]
3895
3896
3897
3898
3899
3900

                # upcast in fp32 if any
                target_dtype = dtype
                if (
                    keep_in_fp32_modules is not None
                    and dtype == torch.float16
3901
3902
3903
                    and any(
                        module_to_keep_in_fp32 in key.split(".") for module_to_keep_in_fp32 in keep_in_fp32_modules
                    )
3904
3905
3906
                ):
                    target_dtype = torch.float32

3907
                if param.device == torch.device("meta"):
3908
                    value = torch.empty(*param.size(), dtype=target_dtype)
3909
                    if (
3910
                        not is_quantized
3911
3912
3913
3914
                        or getattr(hf_quantizer, "requires_parameters_quantization", False)
                        or not hf_quantizer.check_quantized_param(
                            model, param_value=value, param_name=key, state_dict={}
                        )
3915
3916
                    ):
                        set_module_tensor_to_device(model, key, "cpu", value)
3917
                    else:
3918
                        hf_quantizer.create_quantized_param(model, value, key, "cpu", state_dict, unexpected_keys)
3919

3920
        # retrieve uninitialized modules and initialize before maybe overriding that with the pretrained weights.
3921
        if _fast_init:
3922
3923
3924
3925
3926
3927
3928
            if not ignore_mismatched_sizes:
                if remove_prefix_from_model:
                    _loaded_keys = [f"{prefix}.{k}" for k in loaded_keys]
                elif add_prefix_to_model:
                    _loaded_keys = [k[len(prefix) + 1 :] for k in loaded_keys]
                else:
                    _loaded_keys = loaded_keys
3929
                not_initialized_submodules = set_initialized_submodules(model, _loaded_keys)
3930
                # If we're about to tie the output embeds to the input embeds we don't need to init them
3931
3932
3933
                if hasattr(model.config, "tie_word_embeddings") and model.config.tie_word_embeddings:
                    output_embeddings = model.get_output_embeddings()
                    if output_embeddings is not None:
3934
3935
3936
                        # Still need to initialize if there is a bias term since biases are not tied.
                        if not hasattr(output_embeddings, "bias") or output_embeddings.bias is None:
                            output_embeddings._is_hf_initialized = True
3937
3938
            else:
                not_initialized_submodules = dict(model.named_modules())
3939
            # This will only initialize submodules that are not marked as initialized by the line above.
3940
            if is_deepspeed_zero3_enabled() and not is_quantized:
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
                import deepspeed

                not_initialized_parameters = list(
                    set(
                        itertools.chain.from_iterable(
                            submodule.parameters(recurse=False) for submodule in not_initialized_submodules.values()
                        )
                    )
                )
                with deepspeed.zero.GatheredParameters(not_initialized_parameters, modifier_rank=0):
                    model.apply(model._initialize_weights)
            else:
                model.apply(model._initialize_weights)
3954

3955
3956
3957
        # Set some modules to fp32 if any
        if keep_in_fp32_modules is not None:
            for name, param in model.named_parameters():
3958
                if any(module_to_keep_in_fp32 in name.split(".") for module_to_keep_in_fp32 in keep_in_fp32_modules):
3959
3960
                    # param = param.to(torch.float32) does not work here as only in the local scope.
                    param.data = param.data.to(torch.float32)
3961

3962
3963
3964
        # Make sure we are able to load base models as well as derived models (with heads)
        start_prefix = ""
        model_to_load = model
3965
        if len(cls.base_model_prefix) > 0 and not hasattr(model, cls.base_model_prefix) and has_prefix_module:
3966
            start_prefix = cls.base_model_prefix + "."
3967
        if len(cls.base_model_prefix) > 0 and hasattr(model, cls.base_model_prefix) and not has_prefix_module:
3968
            model_to_load = getattr(model, cls.base_model_prefix)
Sylvain Gugger's avatar
Sylvain Gugger committed
3969
3970
            base_model_expected_keys = list(model_to_load.state_dict().keys())
            if any(key in expected_keys_not_prefixed and key not in base_model_expected_keys for key in loaded_keys):
3971
                raise ValueError(
3972
                    "The state dictionary of the model you are trying to load is corrupted. Are you sure it was "
3973
3974
                    "properly saved?"
                )
3975
3976
            if device_map is not None:
                device_map = {k.replace(f"{cls.base_model_prefix}.", ""): v for k, v in device_map.items()}
3977

3978
3979
3980
3981
3982
3983
3984
3985
        def _find_mismatched_keys(
            state_dict,
            model_state_dict,
            loaded_keys,
            add_prefix_to_model,
            remove_prefix_from_model,
            ignore_mismatched_sizes,
        ):
Sylvain Gugger's avatar
Sylvain Gugger committed
3986
3987
3988
            mismatched_keys = []
            if ignore_mismatched_sizes:
                for checkpoint_key in loaded_keys:
3989
3990
3991
                    # If the checkpoint is sharded, we may not have the key here.
                    if checkpoint_key not in state_dict:
                        continue
Sylvain Gugger's avatar
Sylvain Gugger committed
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
                    model_key = checkpoint_key
                    if remove_prefix_from_model:
                        # The model key starts with `prefix` but `checkpoint_key` doesn't so we add it.
                        model_key = f"{prefix}.{checkpoint_key}"
                    elif add_prefix_to_model:
                        # The model key doesn't start with `prefix` but `checkpoint_key` does so we remove it.
                        model_key = ".".join(checkpoint_key.split(".")[1:])

                    if (
                        model_key in model_state_dict
                        and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape
                    ):
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
                        if (
                            state_dict[checkpoint_key].shape[-1] == 1
                            and state_dict[checkpoint_key].numel() * 2 == model_state_dict[model_key].numel()
                        ):
                            # This skips size mismatches for 4-bit weights. Two 4-bit values share an 8-bit container, causing size differences.
                            # Without matching with module type or paramter type it seems like a practical way to detect valid 4bit weights.
                            pass
                        else:
                            mismatched_keys.append(
                                (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
                            )
                            del state_dict[checkpoint_key]
4016
4017
            return mismatched_keys

4018
4019
4020
4021
        if resolved_archive_file is not None:
            folder = os.path.sep.join(resolved_archive_file[0].split(os.path.sep)[:-1])
        else:
            folder = None
Sylvain Gugger's avatar
Sylvain Gugger committed
4022
        if device_map is not None and is_safetensors:
4023
            param_device_map = expand_device_map(device_map, original_loaded_keys, start_prefix)
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
            str_dtype = str(dtype).replace("torch.", "") if dtype is not None else "float32"
            if sharded_metadata is None:
                archive_file = (
                    resolved_archive_file[0]
                    if isinstance(resolved_archive_file, (list, tuple))
                    else resolved_archive_file
                )
                weight_map = {p: archive_file for p in original_loaded_keys}
            else:
                weight_map = {p: os.path.join(folder, f) for p, f in sharded_metadata["weight_map"].items()}
Sylvain Gugger's avatar
Sylvain Gugger committed
4034
            offload_index = {
4035
                p[len(start_prefix) :]: {"safetensors_file": f, "weight_name": p, "dtype": str_dtype}
4036
                for p, f in weight_map.items()
4037
                if p.startswith(start_prefix) and param_device_map[p[len(start_prefix) :]] == "disk"
Sylvain Gugger's avatar
Sylvain Gugger committed
4038
4039
            }

4040
4041
4042
4043
4044
        if state_dict is not None:
            # Whole checkpoint
            mismatched_keys = _find_mismatched_keys(
                state_dict,
                model_state_dict,
4045
                original_loaded_keys,
4046
4047
4048
4049
                add_prefix_to_model,
                remove_prefix_from_model,
                ignore_mismatched_sizes,
            )
Sylvain Gugger's avatar
Sylvain Gugger committed
4050
            error_msgs = _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
Sylvain Gugger's avatar
Sylvain Gugger committed
4051
            offload_index = None
Sylvain Gugger's avatar
Sylvain Gugger committed
4052
        else:
4053
4054
            # Sharded checkpoint or whole but low_cpu_mem_usage==True

Sylvain Gugger's avatar
Sylvain Gugger committed
4055
4056
4057
4058
4059
            # This should always be a list but, just to be sure.
            if not isinstance(resolved_archive_file, list):
                resolved_archive_file = [resolved_archive_file]

            error_msgs = []
4060
            mismatched_keys = []
Sylvain Gugger's avatar
Sylvain Gugger committed
4061
4062
            if not is_safetensors:
                offload_index = {} if device_map is not None and "disk" in device_map.values() else None
4063
4064
4065
4066
4067
4068
4069
            if offload_state_dict:
                state_dict_folder = tempfile.mkdtemp()
                state_dict_index = {}
            else:
                state_dict_folder = None
                state_dict_index = None

4070
            if is_sharded_safetensors:
4071
4072
4073
                disk_only_shard_files = get_disk_only_shard_files(
                    device_map, sharded_metadata=sharded_metadata, start_prefix=start_prefix
                )
Sylvain Gugger's avatar
Sylvain Gugger committed
4074
4075
4076
4077
                disk_only_shard_files = [os.path.join(folder, f) for f in disk_only_shard_files]
            else:
                disk_only_shard_files = []

4078
4079
            if len(resolved_archive_file) > 1:
                resolved_archive_file = logging.tqdm(resolved_archive_file, desc="Loading checkpoint shards")
Sylvain Gugger's avatar
Sylvain Gugger committed
4080
            for shard_file in resolved_archive_file:
Sylvain Gugger's avatar
Sylvain Gugger committed
4081
4082
4083
                # Skip the load for shards that only contain disk-offloaded weights when using safetensors for the offload.
                if shard_file in disk_only_shard_files:
                    continue
4084
                state_dict = load_state_dict(shard_file, is_quantized=is_quantized)
4085

Sylvain Gugger's avatar
Sylvain Gugger committed
4086
4087
                # Mistmatched keys contains tuples key/shape1/shape2 of weights in the checkpoint that have a shape not
                # matching the weights in the model.
4088
4089
4090
                mismatched_keys += _find_mismatched_keys(
                    state_dict,
                    model_state_dict,
4091
                    original_loaded_keys,
4092
4093
4094
4095
                    add_prefix_to_model,
                    remove_prefix_from_model,
                    ignore_mismatched_sizes,
                )
4096
                if low_cpu_mem_usage:
4097
                    if is_fsdp_enabled() and not is_local_dist_rank_0() and not is_quantized:
4098
4099
                        for key, param in model_to_load.state_dict().items():
                            if param.device == torch.device("meta"):
4100
4101
4102
                                set_module_tensor_to_device(
                                    model_to_load, key, "cpu", torch.empty(*param.size(), dtype=dtype)
                                )
4103
                    else:
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
                        new_error_msgs, offload_index, state_dict_index = _load_state_dict_into_meta_model(
                            model_to_load,
                            state_dict,
                            loaded_keys,
                            start_prefix,
                            expected_keys,
                            device_map=device_map,
                            offload_folder=offload_folder,
                            offload_index=offload_index,
                            state_dict_folder=state_dict_folder,
                            state_dict_index=state_dict_index,
                            dtype=dtype,
4116
                            hf_quantizer=hf_quantizer,
4117
4118
                            is_safetensors=is_safetensors,
                            keep_in_fp32_modules=keep_in_fp32_modules,
4119
                            unexpected_keys=unexpected_keys,
4120
4121
                        )
                        error_msgs += new_error_msgs
4122
4123
                else:
                    error_msgs += _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
4124

4125
4126
4127
4128
                # force memory release
                del state_dict
                gc.collect()

4129
            if offload_index is not None and len(offload_index) > 0:
Sylvain Gugger's avatar
Sylvain Gugger committed
4130
4131
4132
                if model != model_to_load:
                    # We need to add the prefix of the base model
                    prefix = cls.base_model_prefix
Sylvain Gugger's avatar
Sylvain Gugger committed
4133
4134
4135
4136
4137
4138
                    if not is_safetensors:
                        for weight_name in offload_index:
                            shutil.move(
                                os.path.join(offload_folder, f"{weight_name}.dat"),
                                os.path.join(offload_folder, f"{prefix}.{weight_name}.dat"),
                            )
Sylvain Gugger's avatar
Sylvain Gugger committed
4139
                    offload_index = {f"{prefix}.{key}": value for key, value in offload_index.items()}
Sylvain Gugger's avatar
Sylvain Gugger committed
4140
4141
4142
                if not is_safetensors:
                    save_offload_index(offload_index, offload_folder)
                    offload_index = None
4143
4144
4145

            if offload_state_dict:
                # Load back temporarily offloaded state dict
4146
                load_offloaded_weights(model_to_load, state_dict_index, state_dict_folder)
4147
4148
                shutil.rmtree(state_dict_folder)

4149
4150
        if len(error_msgs) > 0:
            error_msg = "\n\t".join(error_msgs)
4151
4152
4153
4154
            if "size mismatch" in error_msg:
                error_msg += (
                    "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
                )
4155
4156
            raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")

4157
        if len(unexpected_keys) > 0:
Sylvain Gugger's avatar
Sylvain Gugger committed
4158
            archs = [] if model.config.architectures is None else model.config.architectures
4159
            warner = logger.warning if model.__class__.__name__ in archs else logger.info
Sylvain Gugger's avatar
Sylvain Gugger committed
4160
            warner(
Sylvain Gugger's avatar
Sylvain Gugger committed
4161
4162
4163
4164
4165
4166
4167
                f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when"
                f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are"
                f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task or"
                " with another architecture (e.g. initializing a BertForSequenceClassification model from a"
                " BertForPreTraining model).\n- This IS NOT expected if you are initializing"
                f" {model.__class__.__name__} from the checkpoint of a model that you expect to be exactly identical"
                " (initializing a BertForSequenceClassification model from a BertForSequenceClassification model)."
4168
4169
4170
4171
4172
            )
        else:
            logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
        if len(missing_keys) > 0:
            logger.warning(
Sylvain Gugger's avatar
Sylvain Gugger committed
4173
4174
4175
                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."
4176
            )
4177
        elif len(mismatched_keys) == 0:
4178
            logger.info(
Sylvain Gugger's avatar
Sylvain Gugger committed
4179
4180
4181
4182
                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 checkpoint"
                f" was trained on, you can already use {model.__class__.__name__} for predictions without further"
                " training."
4183
            )
4184
4185
4186
4187
4188
4189
4190
4191
        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(
Sylvain Gugger's avatar
Sylvain Gugger committed
4192
4193
4194
4195
                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."
4196
            )
4197

Sylvain Gugger's avatar
Sylvain Gugger committed
4198
        return model, missing_keys, unexpected_keys, mismatched_keys, offload_index, error_msgs
4199
4200

    def retrieve_modules_from_names(self, names, add_prefix=False, remove_prefix=False):
4201
        module_keys = {".".join(key.split(".")[:-1]) for key in names}
4202

Patrick von Platen's avatar
Patrick von Platen committed
4203
4204
        # torch.nn.ParameterList is a special case where two parameter keywords
        # are appended to the module name, *e.g.* bert.special_embeddings.0
4205
        module_keys = module_keys.union(
4206
            {".".join(key.split(".")[:-2]) for key in names if len(key) > 0 and key[-1].isdigit()}
4207
        )
Patrick von Platen's avatar
Patrick von Platen committed
4208

4209
4210
4211
4212
        retrieved_modules = []
        # retrieve all modules that has at least one missing weight name
        for name, module in self.named_modules():
            if remove_prefix:
4213
4214
                _prefix = f"{self.base_model_prefix}."
                name = name[len(_prefix) :] if name.startswith(_prefix) else name
4215
            elif add_prefix:
Patrick von Platen's avatar
Patrick von Platen committed
4216
                name = ".".join([self.base_model_prefix, name]) if len(name) > 0 else self.base_model_prefix
4217
4218
4219
4220
4221
4222

            if name in module_keys:
                retrieved_modules.append(module)

        return retrieved_modules

4223
    @staticmethod
4224
4225
4226
    def _load_pretrained_model_low_mem(
        model, loaded_state_dict_keys, resolved_archive_file, start_prefix="", hf_quantizer=None
    ):
4227
4228
4229
        """
        This is an experimental function that loads the model using ~1.x model size CPU memory

4230
        Before you call it do:
4231

4232
        1. save which state_dict keys are available
4233
4234
4235
4236
4237
4238
4239
4240
        2. drop state_dict before model is created, since the latter takes 1x model size memory

        Here then we continue:

        3. switch to the meta device all params/buffers that are going to be replaced from the loaded state_dict
        4. load state_dict 2nd time
        5. replace the params/buffers from the state_dict

4241
4242
        Currently, it doesn't handle missing_keys, unexpected_keys, mismatched_keys. It can't handle deepspeed. To
        handle bitsandbytes, needs non-empty hf_quantizer argument.
4243
4244
        """

4245
4246
        _move_model_to_meta(model, loaded_state_dict_keys, start_prefix)
        state_dict = load_state_dict(resolved_archive_file)
4247
4248
4249
4250
4251
4252
4253
4254
4255
        expected_keys = loaded_state_dict_keys  # plug for missing expected_keys. TODO: replace with proper keys
        error_msgs = _load_state_dict_into_meta_model(
            model,
            state_dict,
            loaded_state_dict_keys,
            start_prefix,
            expected_keys=expected_keys,
            hf_quantizer=hf_quantizer,
        )
4256
        return error_msgs
4257

4258
4259
4260
4261
4262
4263
    @classmethod
    def register_for_auto_class(cls, auto_class="AutoModel"):
        """
        Register this class with a given auto class. This should only be used for custom models as the ones in the
        library are already mapped with an auto class.

4264
4265
4266
4267
4268
4269
        <Tip warning={true}>

        This API is experimental and may have some slight breaking changes in the next releases.

        </Tip>

4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
        Args:
            auto_class (`str` or `type`, *optional*, defaults to `"AutoModel"`):
                The auto class to register this new model with.
        """
        if not isinstance(auto_class, str):
            auto_class = auto_class.__name__

        import transformers.models.auto as auto_module

        if not hasattr(auto_module, auto_class):
            raise ValueError(f"{auto_class} is not a valid auto class.")

        cls._auto_class = auto_class

4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
    def to_bettertransformer(self) -> "PreTrainedModel":
        """
        Converts the model to use [PyTorch's native attention
        implementation](https://pytorch.org/docs/stable/generated/torch.nn.MultiheadAttention.html), integrated to
        Transformers through [Optimum library](https://huggingface.co/docs/optimum/bettertransformer/overview). Only a
        subset of all Transformers models are supported.

        PyTorch's attention fastpath allows to speed up inference through kernel fusions and the use of [nested
        tensors](https://pytorch.org/docs/stable/nested.html). Detailed benchmarks can be found in [this blog
        post](https://medium.com/pytorch/bettertransformer-out-of-the-box-performance-for-huggingface-transformers-3fbe27d50ab2).

        Returns:
            [`PreTrainedModel`]: The model converted to BetterTransformer.
        """
        if not is_optimum_available():
            raise ImportError("The package `optimum` is required to use Better Transformer.")

        from optimum.version import __version__ as optimum_version

        if version.parse(optimum_version) < version.parse("1.7.0"):
            raise ImportError(
                f"Please install optimum>=1.7.0 to use Better Transformer. The version {optimum_version} was found."
            )

        from optimum.bettertransformer import BetterTransformer

        return BetterTransformer.transform(self)

    def reverse_bettertransformer(self):
        """
        Reverts the transformation from [`~PreTrainedModel.to_bettertransformer`] so that the original modeling is
        used, for example in order to save the model.

        Returns:
            [`PreTrainedModel`]: The model converted back to the original modeling.
        """
        if not is_optimum_available():
            raise ImportError("The package `optimum` is required to use Better Transformer.")

        from optimum.version import __version__ as optimum_version

        if version.parse(optimum_version) < version.parse("1.7.0"):
            raise ImportError(
                f"Please install optimum>=1.7.0 to use Better Transformer. The version {optimum_version} was found."
            )

        from optimum.bettertransformer import BetterTransformer

        return BetterTransformer.reverse(self)

4334
4335
4336
4337
    def warn_if_padding_and_no_attention_mask(self, input_ids, attention_mask):
        """
        Shows a one-time warning if the input_ids appear to contain padding and no attention mask was given.
        """
4338
4339

        # Skip the check during tracing.
4340
        if is_torch_fx_proxy(input_ids) or torch.jit.is_tracing() or is_torchdynamo_compiling():
4341
4342
            return

4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
        if (attention_mask is not None) or (self.config.pad_token_id is None):
            return

        # Check only the first and last input IDs to reduce overhead.
        if self.config.pad_token_id in input_ids[:, [-1, 0]]:
            warn_string = (
                "We strongly recommend passing in an `attention_mask` since your input_ids may be padded. See "
                "https://huggingface.co/docs/transformers/troubleshooting"
                "#incorrect-output-when-padding-tokens-arent-masked."
            )

            # If the pad token is equal to either BOS, EOS, or SEP, we do not know whether the user should use an
            # attention_mask or not. In this case, we should still show a warning because this is a rare case.
            if (
                (self.config.bos_token_id is not None and self.config.bos_token_id == self.config.pad_token_id)
                or (self.config.eos_token_id is not None and self.config.eos_token_id == self.config.pad_token_id)
                or (self.config.sep_token_id is not None and self.config.sep_token_id == self.config.pad_token_id)
            ):
                warn_string += (
                    f"\nYou may ignore this warning if your `pad_token_id` ({self.config.pad_token_id}) is identical "
                    f"to the `bos_token_id` ({self.config.bos_token_id}), `eos_token_id` ({self.config.eos_token_id}), "
                    f"or the `sep_token_id` ({self.config.sep_token_id}), and your input is not padded."
                )

            logger.warning_once(warn_string)

4369
4370
    @property
    def _is_quantized_training_enabled(self):
4371
        warnings.warn(
4372
4373
4374
4375
4376
4377
4378
4379
4380
            "`_is_quantized_training_enabled` is going to be deprecated in transformers 4.39.0. Please use `model.hf_quantizer.is_trainable` instead",
            FutureWarning,
        )

        if not hasattr(self, "hf_quantizer"):
            return False

        return self.hf_quantizer.is_trainable

thomwolf's avatar
thomwolf committed
4381

4382
PreTrainedModel.push_to_hub = copy_func(PreTrainedModel.push_to_hub)
4383
4384
4385
4386
if PreTrainedModel.push_to_hub.__doc__ is not None:
    PreTrainedModel.push_to_hub.__doc__ = PreTrainedModel.push_to_hub.__doc__.format(
        object="model", object_class="AutoModel", object_files="model file"
    )
4387
4388


thomwolf's avatar
thomwolf committed
4389
class PoolerStartLogits(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4390
4391
    """
    Compute SQuAD start logits from sequence hidden states.
4392

Sylvain Gugger's avatar
Sylvain Gugger committed
4393
    Args:
4394
4395
        config ([`PretrainedConfig`]):
            The config used by the model, will be used to grab the `hidden_size` of the model.
Sylvain Gugger's avatar
Sylvain Gugger committed
4396
4397
4398
    """

    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
4399
        super().__init__()
thomwolf's avatar
thomwolf committed
4400
4401
        self.dense = nn.Linear(config.hidden_size, 1)

Sylvain Gugger's avatar
Sylvain Gugger committed
4402
4403
4404
4405
4406
    def forward(
        self, hidden_states: torch.FloatTensor, p_mask: Optional[torch.FloatTensor] = None
    ) -> torch.FloatTensor:
        """
        Args:
4407
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4408
                The final hidden states of the model.
4409
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4410
4411
                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
                should be masked.
Sylvain Gugger's avatar
Sylvain Gugger committed
4412
4413

        Returns:
4414
            `torch.FloatTensor`: The start logits for SQuAD.
thomwolf's avatar
thomwolf committed
4415
        """
thomwolf's avatar
thomwolf committed
4416
4417
4418
        x = self.dense(hidden_states).squeeze(-1)

        if p_mask is not None:
Lysandre Debut's avatar
Lysandre Debut committed
4419
            if get_parameter_dtype(self) == torch.float16:
4420
4421
4422
                x = x * (1 - p_mask) - 65500 * p_mask
            else:
                x = x * (1 - p_mask) - 1e30 * p_mask
thomwolf's avatar
thomwolf committed
4423
4424
4425
4426
4427
4428

        return x


class PoolerEndLogits(nn.Module):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
4429
    Compute SQuAD end logits from sequence hidden states.
4430

Sylvain Gugger's avatar
Sylvain Gugger committed
4431
    Args:
4432
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
4433
4434
            The config used by the model, will be used to grab the `hidden_size` of the model and the `layer_norm_eps`
            to use.
Sylvain Gugger's avatar
Sylvain Gugger committed
4435
4436
4437
    """

    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
4438
        super().__init__()
thomwolf's avatar
thomwolf committed
4439
4440
4441
4442
4443
        self.dense_0 = nn.Linear(config.hidden_size * 2, config.hidden_size)
        self.activation = nn.Tanh()
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dense_1 = nn.Linear(config.hidden_size, 1)

Sylvain Gugger's avatar
Sylvain Gugger committed
4444
4445
4446
4447
4448
4449
4450
4451
4452
    def forward(
        self,
        hidden_states: torch.FloatTensor,
        start_states: Optional[torch.FloatTensor] = None,
        start_positions: Optional[torch.LongTensor] = None,
        p_mask: Optional[torch.FloatTensor] = None,
    ) -> torch.FloatTensor:
        """
        Args:
4453
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4454
                The final hidden states of the model.
4455
            start_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4456
                The hidden states of the first tokens for the labeled span.
4457
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4458
                The position of the first token for the labeled span.
4459
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4460
4461
                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
                should be masked.
Sylvain Gugger's avatar
Sylvain Gugger committed
4462

4463
        <Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
4464

Stas Bekman's avatar
Stas Bekman committed
4465
4466
        One of `start_states` or `start_positions` should be not `None`. If both are set, `start_positions` overrides
        `start_states`.
4467
4468

        </Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
4469
4470

        Returns:
4471
            `torch.FloatTensor`: The end logits for SQuAD.
thomwolf's avatar
thomwolf committed
4472
        """
4473
4474
4475
        assert (
            start_states is not None or start_positions is not None
        ), "One of start_states, start_positions should be not None"
thomwolf's avatar
thomwolf committed
4476
        if start_positions is not None:
4477
            slen, hsz = hidden_states.shape[-2:]
4478
4479
4480
            start_positions = start_positions[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
            start_states = hidden_states.gather(-2, start_positions)  # shape (bsz, 1, hsz)
            start_states = start_states.expand(-1, slen, -1)  # shape (bsz, slen, hsz)
thomwolf's avatar
thomwolf committed
4481
4482
4483
4484
4485
4486
4487

        x = self.dense_0(torch.cat([hidden_states, start_states], dim=-1))
        x = self.activation(x)
        x = self.LayerNorm(x)
        x = self.dense_1(x).squeeze(-1)

        if p_mask is not None:
Lysandre Debut's avatar
Lysandre Debut committed
4488
            if get_parameter_dtype(self) == torch.float16:
4489
4490
4491
                x = x * (1 - p_mask) - 65500 * p_mask
            else:
                x = x * (1 - p_mask) - 1e30 * p_mask
thomwolf's avatar
thomwolf committed
4492
4493
4494
4495
4496

        return x


class PoolerAnswerClass(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4497
4498
4499
4500
    """
    Compute SQuAD 2.0 answer class from classification and start tokens hidden states.

    Args:
4501
4502
        config ([`PretrainedConfig`]):
            The config used by the model, will be used to grab the `hidden_size` of the model.
Sylvain Gugger's avatar
Sylvain Gugger committed
4503
    """
4504

thomwolf's avatar
thomwolf committed
4505
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
4506
        super().__init__()
thomwolf's avatar
thomwolf committed
4507
4508
4509
4510
        self.dense_0 = nn.Linear(config.hidden_size * 2, config.hidden_size)
        self.activation = nn.Tanh()
        self.dense_1 = nn.Linear(config.hidden_size, 1, bias=False)

Sylvain Gugger's avatar
Sylvain Gugger committed
4511
4512
4513
4514
4515
4516
4517
    def forward(
        self,
        hidden_states: torch.FloatTensor,
        start_states: Optional[torch.FloatTensor] = None,
        start_positions: Optional[torch.LongTensor] = None,
        cls_index: Optional[torch.LongTensor] = None,
    ) -> torch.FloatTensor:
4518
4519
        """
        Args:
4520
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4521
                The final hidden states of the model.
4522
            start_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4523
                The hidden states of the first tokens for the labeled span.
4524
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4525
                The position of the first token for the labeled span.
4526
4527
4528
4529
            cls_index (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
                Position of the CLS token for each sentence in the batch. If `None`, takes the last token.

        <Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
4530

Stas Bekman's avatar
Stas Bekman committed
4531
4532
        One of `start_states` or `start_positions` should be not `None`. If both are set, `start_positions` overrides
        `start_states`.
Sylvain Gugger's avatar
Sylvain Gugger committed
4533

4534
        </Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
4535
4536

        Returns:
4537
            `torch.FloatTensor`: The SQuAD 2.0 answer class.
thomwolf's avatar
thomwolf committed
4538
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
4539
        # No dependency on end_feature so that we can obtain one single `cls_logits` for each sample.
4540
        hsz = hidden_states.shape[-1]
4541
4542
4543
        assert (
            start_states is not None or start_positions is not None
        ), "One of start_states, start_positions should be not None"
thomwolf's avatar
thomwolf committed
4544
        if start_positions is not None:
4545
4546
            start_positions = start_positions[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
            start_states = hidden_states.gather(-2, start_positions).squeeze(-2)  # shape (bsz, hsz)
thomwolf's avatar
thomwolf committed
4547
4548

        if cls_index is not None:
4549
4550
            cls_index = cls_index[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
            cls_token_state = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, hsz)
thomwolf's avatar
thomwolf committed
4551
        else:
4552
            cls_token_state = hidden_states[:, -1, :]  # shape (bsz, hsz)
thomwolf's avatar
thomwolf committed
4553
4554
4555
4556
4557
4558
4559
4560

        x = self.dense_0(torch.cat([start_states, cls_token_state], dim=-1))
        x = self.activation(x)
        x = self.dense_1(x).squeeze(-1)

        return x


4561
4562
4563
@dataclass
class SquadHeadOutput(ModelOutput):
    """
4564
    Base class for outputs of question answering models using a [`~modeling_utils.SQuADHead`].
4565
4566

    Args:
4567
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned if both `start_positions` and `end_positions` are provided):
Sylvain Gugger's avatar
Sylvain Gugger committed
4568
4569
            Classification loss as the sum of start token, end token (and is_impossible if provided) classification
            losses.
4570
        start_top_log_probs (`torch.FloatTensor` of shape `(batch_size, config.start_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
4571
            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
4572
        start_top_index (`torch.LongTensor` of shape `(batch_size, config.start_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
4573
            Indices for the top config.start_n_top start token possibilities (beam-search).
4574
4575
        end_top_log_probs (`torch.FloatTensor` of shape `(batch_size, config.start_n_top * config.end_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
            Log probabilities for the top `config.start_n_top * config.end_n_top` end token possibilities
Sylvain Gugger's avatar
Sylvain Gugger committed
4576
            (beam-search).
4577
4578
4579
4580
        end_top_index (`torch.LongTensor` of shape `(batch_size, config.start_n_top * config.end_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
            Indices for the top `config.start_n_top * config.end_n_top` end token possibilities (beam-search).
        cls_logits (`torch.FloatTensor` of shape `(batch_size,)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
            Log probabilities for the `is_impossible` label of the answers.
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591

    """

    loss: Optional[torch.FloatTensor] = None
    start_top_log_probs: Optional[torch.FloatTensor] = None
    start_top_index: Optional[torch.LongTensor] = None
    end_top_log_probs: Optional[torch.FloatTensor] = None
    end_top_index: Optional[torch.LongTensor] = None
    cls_logits: Optional[torch.FloatTensor] = None


thomwolf's avatar
thomwolf committed
4592
class SQuADHead(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4593
4594
    r"""
    A SQuAD head inspired by XLNet.
4595

Sylvain Gugger's avatar
Sylvain Gugger committed
4596
    Args:
4597
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
4598
4599
            The config used by the model, will be used to grab the `hidden_size` of the model and the `layer_norm_eps`
            to use.
thomwolf's avatar
thomwolf committed
4600
    """
4601

thomwolf's avatar
thomwolf committed
4602
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
4603
        super().__init__()
thomwolf's avatar
thomwolf committed
4604
4605
4606
4607
4608
4609
4610
        self.start_n_top = config.start_n_top
        self.end_n_top = config.end_n_top

        self.start_logits = PoolerStartLogits(config)
        self.end_logits = PoolerEndLogits(config)
        self.answer_class = PoolerAnswerClass(config)

Sylvain Gugger's avatar
Sylvain Gugger committed
4611
    @replace_return_docstrings(output_type=SquadHeadOutput, config_class=PretrainedConfig)
4612
    def forward(
4613
        self,
Sylvain Gugger's avatar
Sylvain Gugger committed
4614
4615
4616
4617
4618
4619
        hidden_states: torch.FloatTensor,
        start_positions: Optional[torch.LongTensor] = None,
        end_positions: Optional[torch.LongTensor] = None,
        cls_index: Optional[torch.LongTensor] = None,
        is_impossible: Optional[torch.LongTensor] = None,
        p_mask: Optional[torch.FloatTensor] = None,
4620
        return_dict: bool = False,
Sylvain Gugger's avatar
Sylvain Gugger committed
4621
4622
    ) -> Union[SquadHeadOutput, Tuple[torch.FloatTensor]]:
        """
Lysandre's avatar
Lysandre committed
4623
        Args:
4624
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Lysandre's avatar
Lysandre committed
4625
                Final hidden states of the model on the sequence tokens.
4626
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Lysandre's avatar
Lysandre committed
4627
                Positions of the first token for the labeled span.
4628
            end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Lysandre's avatar
Lysandre committed
4629
                Positions of the last token for the labeled span.
4630
4631
4632
            cls_index (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
                Position of the CLS token for each sentence in the batch. If `None`, takes the last token.
            is_impossible (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Lysandre's avatar
Lysandre committed
4633
                Whether the question has a possible answer in the paragraph or not.
4634
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4635
4636
                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
                should be masked.
4637
            return_dict (`bool`, *optional*, defaults to `False`):
4638
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
Sylvain Gugger's avatar
Sylvain Gugger committed
4639

Lysandre's avatar
Lysandre committed
4640
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
4641
        """
thomwolf's avatar
thomwolf committed
4642
        start_logits = self.start_logits(hidden_states, p_mask=p_mask)
thomwolf's avatar
thomwolf committed
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665

        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, let's remove the dimension added by batch splitting
            for x in (start_positions, end_positions, cls_index, is_impossible):
                if x is not None and x.dim() > 1:
                    x.squeeze_(-1)

            # during training, compute the end logits based on the ground truth of the start position
            end_logits = self.end_logits(hidden_states, start_positions=start_positions, p_mask=p_mask)

            loss_fct = CrossEntropyLoss()
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

            if cls_index is not None and is_impossible is not None:
                # Predict answerability from the representation of CLS and START
                cls_logits = self.answer_class(hidden_states, start_positions=start_positions, cls_index=cls_index)
                loss_fct_cls = nn.BCEWithLogitsLoss()
                cls_loss = loss_fct_cls(cls_logits, is_impossible)

                # note(zhiliny): by default multiply the loss by 0.5 so that the scale is comparable to start_loss and end_loss
                total_loss += cls_loss * 0.5
4666

4667
            return SquadHeadOutput(loss=total_loss) if return_dict else (total_loss,)
thomwolf's avatar
thomwolf committed
4668
4669
4670
4671

        else:
            # during inference, compute the end logits based on beam search
            bsz, slen, hsz = hidden_states.size()
4672
            start_log_probs = nn.functional.softmax(start_logits, dim=-1)  # shape (bsz, slen)
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683

            start_top_log_probs, start_top_index = torch.topk(
                start_log_probs, self.start_n_top, dim=-1
            )  # shape (bsz, start_n_top)
            start_top_index_exp = start_top_index.unsqueeze(-1).expand(-1, -1, hsz)  # shape (bsz, start_n_top, hsz)
            start_states = torch.gather(hidden_states, -2, start_top_index_exp)  # shape (bsz, start_n_top, hsz)
            start_states = start_states.unsqueeze(1).expand(-1, slen, -1, -1)  # shape (bsz, slen, start_n_top, hsz)

            hidden_states_expanded = hidden_states.unsqueeze(2).expand_as(
                start_states
            )  # shape (bsz, slen, start_n_top, hsz)
thomwolf's avatar
thomwolf committed
4684
4685
            p_mask = p_mask.unsqueeze(-1) if p_mask is not None else None
            end_logits = self.end_logits(hidden_states_expanded, start_states=start_states, p_mask=p_mask)
4686
            end_log_probs = nn.functional.softmax(end_logits, dim=1)  # shape (bsz, slen, start_n_top)
thomwolf's avatar
thomwolf committed
4687

4688
4689
4690
            end_top_log_probs, end_top_index = torch.topk(
                end_log_probs, self.end_n_top, dim=1
            )  # shape (bsz, end_n_top, start_n_top)
thomwolf's avatar
thomwolf committed
4691
4692
4693
4694
4695
4696
            end_top_log_probs = end_top_log_probs.view(-1, self.start_n_top * self.end_n_top)
            end_top_index = end_top_index.view(-1, self.start_n_top * self.end_n_top)

            start_states = torch.einsum("blh,bl->bh", hidden_states, start_log_probs)
            cls_logits = self.answer_class(hidden_states, start_states=start_states, cls_index=cls_index)

4697
            if not return_dict:
4698
4699
4700
4701
4702
4703
4704
4705
4706
                return (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits)
            else:
                return SquadHeadOutput(
                    start_top_log_probs=start_top_log_probs,
                    start_top_index=start_top_index,
                    end_top_log_probs=end_top_log_probs,
                    end_top_index=end_top_index,
                    cls_logits=cls_logits,
                )
thomwolf's avatar
thomwolf committed
4707
4708
4709


class SequenceSummary(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4710
4711
4712
4713
    r"""
    Compute a single vector summary of a sequence hidden states.

    Args:
4714
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
4715
4716
            The config used by the model. Relevant arguments in the config class of the model are (refer to the actual
            config class of your model for the default values it uses):
Sylvain Gugger's avatar
Sylvain Gugger committed
4717

4718
            - **summary_type** (`str`) -- The method to use to make this summary. Accepted values are:
Sylvain Gugger's avatar
Sylvain Gugger committed
4719

4720
4721
4722
4723
4724
                - `"last"` -- Take the last token hidden state (like XLNet)
                - `"first"` -- Take the first token hidden state (like Bert)
                - `"mean"` -- Take the mean of all tokens hidden states
                - `"cls_index"` -- Supply a Tensor of classification token position (GPT/GPT-2)
                - `"attn"` -- Not implemented now, use multi-head attention
Sylvain Gugger's avatar
Sylvain Gugger committed
4725

4726
            - **summary_use_proj** (`bool`) -- Add a projection after the vector extraction.
Sylvain Gugger's avatar
Sylvain Gugger committed
4727
4728
4729
4730
4731
4732
            - **summary_proj_to_labels** (`bool`) -- If `True`, the projection outputs to `config.num_labels` classes
              (otherwise to `config.hidden_size`).
            - **summary_activation** (`Optional[str]`) -- Set to `"tanh"` to add a tanh activation to the output,
              another string or `None` will add no activation.
            - **summary_first_dropout** (`float`) -- Optional dropout probability before the projection and activation.
            - **summary_last_dropout** (`float`)-- Optional dropout probability after the projection and activation.
thomwolf's avatar
thomwolf committed
4733
    """
4734

4735
    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
4736
        super().__init__()
thomwolf's avatar
thomwolf committed
4737

4738
        self.summary_type = getattr(config, "summary_type", "last")
4739
        if self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
4740
4741
4742
4743
4744
            # We should use a standard multi-head attention module with absolute positional embedding for that.
            # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
            # We can probably just use the multi-head attention module of PyTorch >=1.1.0
            raise NotImplementedError

thomwolf's avatar
thomwolf committed
4745
        self.summary = Identity()
4746
4747
        if hasattr(config, "summary_use_proj") and config.summary_use_proj:
            if hasattr(config, "summary_proj_to_labels") and config.summary_proj_to_labels and config.num_labels > 0:
4748
                num_classes = config.num_labels
thomwolf's avatar
thomwolf committed
4749
4750
4751
4752
            else:
                num_classes = config.hidden_size
            self.summary = nn.Linear(config.hidden_size, num_classes)

4753
        activation_string = getattr(config, "summary_activation", None)
Lysandre's avatar
Lysandre committed
4754
        self.activation: Callable = get_activation(activation_string) if activation_string else Identity()
thomwolf's avatar
thomwolf committed
4755

thomwolf's avatar
thomwolf committed
4756
        self.first_dropout = Identity()
4757
        if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
4758
4759
            self.first_dropout = nn.Dropout(config.summary_first_dropout)

thomwolf's avatar
thomwolf committed
4760
        self.last_dropout = Identity()
4761
        if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
4762
            self.last_dropout = nn.Dropout(config.summary_last_dropout)
thomwolf's avatar
thomwolf committed
4763

Sylvain Gugger's avatar
Sylvain Gugger committed
4764
4765
4766
4767
4768
4769
4770
    def forward(
        self, hidden_states: torch.FloatTensor, cls_index: Optional[torch.LongTensor] = None
    ) -> torch.FloatTensor:
        """
        Compute a single vector summary of a sequence hidden states.

        Args:
4771
            hidden_states (`torch.FloatTensor` of shape `[batch_size, seq_len, hidden_size]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4772
                The hidden states of the last layer.
4773
            cls_index (`torch.LongTensor` of shape `[batch_size]` or `[batch_size, ...]` where ... are optional leading dimensions of `hidden_states`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4774
                Used if `summary_type == "cls_index"` and takes the last token of the sequence as classification token.
Sylvain Gugger's avatar
Sylvain Gugger committed
4775
4776

        Returns:
4777
            `torch.FloatTensor`: The summary of the sequence hidden states.
thomwolf's avatar
thomwolf committed
4778
        """
4779
        if self.summary_type == "last":
thomwolf's avatar
thomwolf committed
4780
            output = hidden_states[:, -1]
4781
        elif self.summary_type == "first":
thomwolf's avatar
thomwolf committed
4782
            output = hidden_states[:, 0]
4783
        elif self.summary_type == "mean":
thomwolf's avatar
thomwolf committed
4784
            output = hidden_states.mean(dim=1)
4785
        elif self.summary_type == "cls_index":
thomwolf's avatar
thomwolf committed
4786
            if cls_index is None:
Lysandre's avatar
Lysandre committed
4787
4788
4789
4790
4791
                cls_index = torch.full_like(
                    hidden_states[..., :1, :],
                    hidden_states.shape[-2] - 1,
                    dtype=torch.long,
                )
thomwolf's avatar
thomwolf committed
4792
            else:
thomwolf's avatar
thomwolf committed
4793
                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
4794
                cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
thomwolf's avatar
thomwolf committed
4795
            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
4796
4797
            output = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, XX, hidden_size)
        elif self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
4798
4799
            raise NotImplementedError

4800
        output = self.first_dropout(output)
thomwolf's avatar
thomwolf committed
4801
4802
        output = self.summary(output)
        output = self.activation(output)
4803
        output = self.last_dropout(output)
thomwolf's avatar
thomwolf committed
4804
4805
4806
4807

        return output


4808
def unwrap_model(model: nn.Module) -> nn.Module:
4809
4810
4811
4812
    """
    Recursively unwraps a model from potential containers (as used in distributed training).

    Args:
4813
        model (`torch.nn.Module`): The model to unwrap.
4814
4815
4816
4817
4818
4819
    """
    # since there could be multiple levels of wrapping, unwrap recursively
    if hasattr(model, "module"):
        return unwrap_model(model.module)
    else:
        return model
Sylvain Gugger's avatar
Sylvain Gugger committed
4820
4821


4822
def expand_device_map(device_map, param_names, start_prefix):
Sylvain Gugger's avatar
Sylvain Gugger committed
4823
4824
4825
4826
    """
    Expand a device map to return the correspondance parameter name to device.
    """
    new_device_map = {}
4827
    param_names = [p[len(start_prefix) :] for p in param_names if p.startswith(start_prefix)]
Sylvain Gugger's avatar
Sylvain Gugger committed
4828
    for module, device in device_map.items():
4829
4830
4831
        new_device_map.update(
            {p: device for p in param_names if p == module or p.startswith(f"{module}.") or module == ""}
        )
Sylvain Gugger's avatar
Sylvain Gugger committed
4832
4833
4834
    return new_device_map


4835
def get_disk_only_shard_files(device_map, sharded_metadata, start_prefix):
Sylvain Gugger's avatar
Sylvain Gugger committed
4836
4837
4838
    """
    Returns the list of shard files containing only weights offloaded to disk.
    """
4839
4840
4841
4842

    weight_map = {
        p[len(start_prefix) :]: v for p, v in sharded_metadata["weight_map"].items() if p.startswith(start_prefix)
    }
Sylvain Gugger's avatar
Sylvain Gugger committed
4843
    files_content = collections.defaultdict(list)
4844
    for weight_name, filename in weight_map.items():
Sylvain Gugger's avatar
Sylvain Gugger committed
4845
4846
4847
4848
4849
        while len(weight_name) > 0 and weight_name not in device_map:
            weight_name = ".".join(weight_name.split(".")[:-1])
        files_content[filename].append(device_map[weight_name])

    return [fname for fname, devices in files_content.items() if set(devices) == {"disk"}]