modeling_utils.py 246 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 huggingface_hub import split_torch_state_dict_into_shards
38
from packaging import version
Sylvain Gugger's avatar
Sylvain Gugger committed
39
from torch import Tensor, nn
40
from torch.nn import CrossEntropyLoss, Identity
41
from torch.utils.checkpoint import checkpoint
42

43
from .activations import get_activation
44
from .configuration_utils import PretrainedConfig
45
from .dynamic_module_utils import custom_object_save
46
from .generation import GenerationConfig, GenerationMixin
47
from .integrations import PeftAdapterMixin, deepspeed_config, is_deepspeed_zero3_enabled
48
49
50
51
from .pytorch_utils import (  # noqa: F401
    Conv1D,
    apply_chunking_to_forward,
    find_pruneable_heads_and_indices,
Thomas Wang's avatar
Thomas Wang committed
52
    id_tensor_storage,
53
    is_torch_greater_or_equal_than_1_13,
54
55
56
57
    prune_conv1d_layer,
    prune_layer,
    prune_linear_layer,
)
58
from .quantizers import AutoHfQuantizer, HfQuantizer
59
from .quantizers.quantizers_utils import get_module_from_name
60
from .safetensors_conversion import auto_conversion
61
from .utils import (
62
63
    ADAPTER_SAFE_WEIGHTS_NAME,
    ADAPTER_WEIGHTS_NAME,
64
    CONFIG_NAME,
Aymeric Augustin's avatar
Aymeric Augustin committed
65
    DUMMY_INPUTS,
66
    FLAX_WEIGHTS_NAME,
67
68
    SAFE_WEIGHTS_INDEX_NAME,
    SAFE_WEIGHTS_NAME,
69
70
    TF2_WEIGHTS_NAME,
    TF_WEIGHTS_NAME,
Sylvain Gugger's avatar
Sylvain Gugger committed
71
    WEIGHTS_INDEX_NAME,
72
    WEIGHTS_NAME,
73
    ContextManagers,
74
    ModelOutput,
Sylvain Gugger's avatar
Sylvain Gugger committed
75
    PushToHubMixin,
76
    cached_file,
77
    copy_func,
78
    download_url,
79
    extract_commit_hash,
80
    has_file,
81
    is_accelerate_available,
82
    is_bitsandbytes_available,
83
    is_flash_attn_2_available,
84
    is_offline_mode,
85
    is_optimum_available,
86
    is_peft_available,
87
    is_remote_url,
88
    is_safetensors_available,
89
    is_torch_sdpa_available,
90
    is_torch_xla_available,
91
    logging,
Sylvain Gugger's avatar
Sylvain Gugger committed
92
    replace_return_docstrings,
93
    strtobool,
94
)
95
from .utils.hub import convert_file_size_to_int, create_and_tag_model_card, get_checkpoint_shard_files
96
97
98
99
100
101
from .utils.import_utils import (
    ENV_VARS_TRUE_VALUES,
    is_sagemaker_mp_enabled,
    is_torch_fx_proxy,
    is_torchdynamo_compiling,
)
102
from .utils.quantization_config import BitsAndBytesConfig, QuantizationMethod
103

Aymeric Augustin's avatar
Aymeric Augustin committed
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
PARAM_RENAME_WARNING = "A parameter name that contains `{}` will be renamed internally to `{}`. Please use a different name to suppress this warning."

109

110
111
if is_accelerate_available():
    from accelerate import dispatch_model, infer_auto_device_map, init_empty_weights
112
    from accelerate.hooks import add_hook_to_module
113
    from accelerate.utils import (
114
        check_tied_parameters_on_same_device,
115
        extract_model_from_parallel,
116
        find_tied_parameters,
117
        get_balanced_memory,
Marc Sun's avatar
Marc Sun committed
118
        get_max_memory,
119
120
121
122
123
124
        load_offloaded_weights,
        offload_weight,
        save_offload_index,
        set_module_tensor_to_device,
    )

125
126
127
128
    accelerate_version = version.parse(importlib.metadata.version("accelerate"))
    if accelerate_version >= version.parse("0.31"):
        from accelerate.utils.modeling import get_state_dict_from_offload

129
130
131
132
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
133

Lysandre Debut's avatar
Lysandre Debut committed
134
logger = logging.get_logger(__name__)
135

136
137
138
139

_init_weights = True


140
def is_fsdp_enabled():
141
142
143
144
    return (
        torch.distributed.is_available()
        and torch.distributed.is_initialized()
        and strtobool(os.environ.get("ACCELERATE_USE_FSDP", "False")) == 1
145
        and strtobool(os.environ.get("FSDP_CPU_RAM_EFFICIENT_LOADING", "False")) == 1
146
    )
147
148


149
150
151
152
153
154
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
    )
155
156


157
158
159
160
161
162
163
164
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

165
166
167
if is_peft_available():
    from .utils import find_adapter_config_file

168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
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,
}

185

186
187
188
189
190
191
192
193
@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
194
    old_init_weights = _init_weights
195

196
197
    if _enable:
        _init_weights = False
198
199
200
201
202
203
204

        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)
205
206
207
    try:
        yield
    finally:
208
        _init_weights = old_init_weights
209
210
211
212
        if _enable:
            # # Restore the original initialization functions
            for name, init_func in TORCH_INIT_FUNCTIONS.items():
                setattr(torch.nn.init, name, init_func)
213
214


Lysandre Debut's avatar
Lysandre Debut committed
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
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


230
231
232
233
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
234
235
236
    try:
        return next(parameter.parameters()).dtype
    except StopIteration:
Sylvain Gugger's avatar
Sylvain Gugger committed
237
        # For nn.DataParallel compatibility in PyTorch > 1.5
Lysandre Debut's avatar
Lysandre Debut committed
238
239
240
241
242
243
244
245
246
247

        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


248
249
def get_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
250
    Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.
251
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
252
253
254
255
    last_dtype = None
    for t in parameter.parameters():
        last_dtype = t.dtype
        if t.is_floating_point():
256
257
258
            # 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
259
260
            # 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():
261
                return torch.bfloat16
262
            if XLA_DOWNCAST_BF16 in ENV_VARS_TRUE_VALUES and is_torch_xla_available():
263
                if t.dtype == torch.float:
264
                    return torch.bfloat16
265
266
                if t.dtype == torch.double:
                    return torch.float32
Sylvain Gugger's avatar
Sylvain Gugger committed
267
            return t.dtype
268

Sylvain Gugger's avatar
Sylvain Gugger committed
269
270
271
    if last_dtype is not None:
        # if no floating dtype was found return whatever the first dtype is
        return last_dtype
272

273
274
275
276
277
278
279
280
281
282
283
284
285
    # 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:
286
287
        # fallback to the last dtype
        return last_tuple[1].dtype
288

289
290
291
292
293
294
295
    # fallback to buffer dtype
    for t in parameter.buffers():
        last_dtype = t.dtype
        if t.is_floating_point():
            return t.dtype
    return last_dtype

296
297
298
299
300
301
302
303
304
305
306
307
308
309

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
310
    Returns the first found floating dtype in `state_dict` if there is one, otherwise returns the first dtype.
311
312
313
314
315
316
317
    """
    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
318
        return next(state_dict.values()).dtype
319
320


Sylvain Gugger's avatar
Sylvain Gugger committed
321
322
323
324
325
326
327
328
329
330
331
332
333
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
334
    bit_search = re.search(r"[^\d](\d+)_?", str(dtype))
Sylvain Gugger's avatar
Sylvain Gugger committed
335
336
337
338
339
340
    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


341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
def check_support_param_buffer_assignment(model_to_load, state_dict, start_prefix=""):
    """
    Checks if `model_to_load` supports param buffer assignment (such
    as when loading in empty weights) by first checking
    if the model explicitly disables it, then by ensuring that the state dict keys
    are a subset of the model's parameters.
    """
    if len([key for key in state_dict if key.startswith(start_prefix)]) == 0:
        return False

    # Some models explicitly do not support param buffer assignment
    if not getattr(model_to_load, "_supports_param_buffer_assignment", False):
        logger.debug(
            f"{model_to_load.__class__.__name__} does not support param buffer assignment, loading will be slower"
        )
        return False

    # If the model does, the incoming `state_dict` and the `model_to_load` must be the same dtype
    first_key = list(model_to_load.state_dict().keys())[0]
    if start_prefix + first_key in state_dict:
        return state_dict[start_prefix + first_key].dtype == model_to_load.state_dict()[first_key].dtype

    # For cases when the `state_dict` doesn't contain real weights to the model (`test_model_weights_reload_no_missing_tied_weights`)
    return False


367
368
369
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
370
371
372
373
374
375
376
377
378
379
380
    """
    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
381
    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
382
383
384
385
386
387
388
389
390
    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"`).
391
392
        weights_name (`str`, *optional*, defaults to `"pytorch_model.bin"`):
            The name of the model save file.
Sylvain Gugger's avatar
Sylvain Gugger committed
393
    """
394
395
396
397
    logger.warning(
        "Note that `shard_checkpoint` is deprecated and will be removed in v4.44. We recommend you using "
        "split_torch_state_dict_into_shards from huggingface_hub library"
    )
Sylvain Gugger's avatar
Sylvain Gugger committed
398
399
    max_shard_size = convert_file_size_to_int(max_shard_size)

Thomas Wang's avatar
Thomas Wang committed
400
401
    sharded_state_dicts = [{}]
    last_block_size = 0
Sylvain Gugger's avatar
Sylvain Gugger committed
402
    total_size = 0
Thomas Wang's avatar
Thomas Wang committed
403
    storage_id_to_block = {}
Sylvain Gugger's avatar
Sylvain Gugger committed
404
405

    for key, weight in state_dict.items():
406
407
408
409
410
411
        # 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
412
413

        # If a `weight` shares the same underlying storage as another tensor, we put `weight` in the same `block`
414
        if storage_id in storage_id_to_block and weight.device != torch.device("meta"):
Thomas Wang's avatar
Thomas Wang committed
415
416
417
418
            block_id = storage_id_to_block[storage_id]
            sharded_state_dicts[block_id][key] = weight
            continue

Sylvain Gugger's avatar
Sylvain Gugger committed
419
        weight_size = weight.numel() * dtype_byte_size(weight.dtype)
Sylvain Gugger's avatar
Sylvain Gugger committed
420
421
422
        # 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
423
424
            sharded_state_dicts.append({})
            last_block_size = 0
Sylvain Gugger's avatar
Sylvain Gugger committed
425

Thomas Wang's avatar
Thomas Wang committed
426
427
        sharded_state_dicts[-1][key] = weight
        last_block_size += weight_size
Sylvain Gugger's avatar
Sylvain Gugger committed
428
        total_size += weight_size
Thomas Wang's avatar
Thomas Wang committed
429
        storage_id_to_block[storage_id] = len(sharded_state_dicts) - 1
Sylvain Gugger's avatar
Sylvain Gugger committed
430
431
432

    # If we only have one shard, we return it
    if len(sharded_state_dicts) == 1:
433
        return {weights_name: sharded_state_dicts[0]}, None
Sylvain Gugger's avatar
Sylvain Gugger committed
434
435
436
437
438

    # Otherwise, let's build the index
    weight_map = {}
    shards = {}
    for idx, shard in enumerate(sharded_state_dicts):
439
440
441
442
        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
443
444
445
446
447
448
449
450
451
452
        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


453
def load_sharded_checkpoint(model, folder, strict=True, prefer_safe=True):
454
455
456
457
458
459
460
461
462
463
464
465
466
    """
    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.
467
468
469
        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.
470
471
472
473
474
475
476
477

    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)
478
    safe_index_file = os.path.join(folder, SAFE_WEIGHTS_INDEX_NAME)
479

480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
    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:
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
        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
523
524
    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)
525

526
    for shard_file in shard_files:
527
        state_dict = loader(os.path.join(folder, shard_file))
528
529
        model.load_state_dict(state_dict, strict=False)

530
        # Make sure memory is freed before we load the next state dict.
531
532
533
534
535
536
537
        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)


538
def load_state_dict(checkpoint_file: Union[str, os.PathLike], is_quantized: bool = False):
Sylvain Gugger's avatar
Sylvain Gugger committed
539
540
541
    """
    Reads a PyTorch checkpoint file, returning properly formatted errors if they arise.
    """
542
543
544
545
    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()
546
        if metadata.get("format") not in ["pt", "tf", "flax", "mlx"]:
547
548
549
550
551
            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
552
    try:
553
        if (
554
555
556
            (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:
557
558
559
            map_location = "meta"
        else:
            map_location = "cpu"
560
561
562
563
564
565
566
567
568
        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
569
        weights_only_kwarg = {"weights_only": True} if is_torch_greater_or_equal_than_1_13 else {}
570
571
572
        return torch.load(
            checkpoint_file,
            map_location=map_location,
Yih-Dar's avatar
Yih-Dar committed
573
            **weights_only_kwarg,
574
575
            **extra_args,
        )
Sylvain Gugger's avatar
Sylvain Gugger committed
576
577
578
    except Exception as e:
        try:
            with open(checkpoint_file) as f:
579
                if f.read(7) == "version":
Sylvain Gugger's avatar
Sylvain Gugger committed
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
                    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."
            )


598
599
600
601
602
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.
    """
603
    not_initialized_submodules = {}
604
    for module_name, module in model.named_modules():
605
606
        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()):
607
            module._is_hf_initialized = True
608
609
610
        else:
            not_initialized_submodules[module_name] = module
    return not_initialized_submodules
611
612


613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
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


686
def _load_state_dict_into_model(model_to_load, state_dict, start_prefix, assign_to_params_buffers=False):
Sylvain Gugger's avatar
Sylvain Gugger committed
687
688
689
690
691
692
    # 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:
693
            logger.warning(PARAM_RENAME_WARNING.format("gamma", "weight"))
Sylvain Gugger's avatar
Sylvain Gugger committed
694
695
            new_key = key.replace("gamma", "weight")
        if "beta" in key:
696
            logger.warning(PARAM_RENAME_WARNING.format("beta", "bias"))
Sylvain Gugger's avatar
Sylvain Gugger committed
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
            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.
714
    def load(module: nn.Module, state_dict, prefix="", assign_to_params_buffers=False):
Sylvain Gugger's avatar
Sylvain Gugger committed
715
        local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
716
717
        local_metadata["assign_to_params_buffers"] = assign_to_params_buffers

718
        args = (state_dict, prefix, local_metadata, True, [], [], error_msgs)
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
        # 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
738
739
740

        for name, child in module._modules.items():
            if child is not None:
741
                load(child, state_dict, prefix + name + ".", assign_to_params_buffers)
Sylvain Gugger's avatar
Sylvain Gugger committed
742

743
    load(model_to_load, state_dict, prefix=start_prefix, assign_to_params_buffers=assign_to_params_buffers)
744
745
746
    # 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
747
748
749
750

    return error_msgs


751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
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)


800
801
802
803
804
805
806
807
808
809
810
811
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,
812
    hf_quantizer=None,
Sylvain Gugger's avatar
Sylvain Gugger committed
813
    is_safetensors=False,
814
    keep_in_fp32_modules=None,
815
    unexpected_keys=None,  # passing `unexpected` for cleanup from quantization items
816
):
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
    """
    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 = []

836
837
    old_keys = []
    new_keys = []
838
    is_quantized = hf_quantizer is not None
839
840
841
    for key in state_dict.keys():
        new_key = None
        if "gamma" in key:
842
            logger.warning(PARAM_RENAME_WARNING.format("gamma", "weight"))
843
844
            new_key = key.replace("gamma", "weight")
        if "beta" in key:
845
            logger.warning(PARAM_RENAME_WARNING.format("beta", "bias"))
846
847
848
849
850
851
            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)
852

853
854
855
856
857
858
859
860
861
    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
862
        set_module_kwargs = {}
863

864
        # We convert floating dtypes to the `dtype` passed. We want to keep the buffers/params
865
866
        # in int/uint/bool and not cast them.
        if dtype is not None and torch.is_floating_point(param):
867
868
            if (
                keep_in_fp32_modules is not None
869
870
871
                and any(
                    module_to_keep_in_fp32 in param_name.split(".") for module_to_keep_in_fp32 in keep_in_fp32_modules
                )
872
873
874
                and dtype == torch.float16
            ):
                param = param.to(torch.float32)
875
876
877
878
879

                # 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
880
881
            else:
                param = param.to(dtype)
882

883
884
885
886
887
888
889
890
891
892
893
894
        # 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:
895
                param = param.to(old_param.dtype)
896

897
898
899
            if old_param.is_contiguous():
                param = param.contiguous()

900
901
        set_module_kwargs["value"] = param

902
903
904
905
906
907
908
909
910
911
912
        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]
913

914
        if param_device == "disk":
Sylvain Gugger's avatar
Sylvain Gugger committed
915
916
            if not is_safetensors:
                offload_index = offload_weight(param, param_name, offload_folder, offload_index)
917
        elif param_device == "cpu" and state_dict_index is not None:
918
            state_dict_index = offload_weight(param, param_name, state_dict_folder, state_dict_index)
919
        elif (
920
            not is_quantized
921
            or (not hf_quantizer.requires_parameters_quantization)
922
923
924
925
926
            or (
                not hf_quantizer.check_quantized_param(
                    model, param, param_name, state_dict, param_device=param_device, device_map=device_map
                )
            )
927
928
        ):
            # For backward compatibility with older versions of `accelerate` and for non-quantized params
929
            set_module_tensor_to_device(model, param_name, param_device, **set_module_kwargs)
930
        else:
931
            hf_quantizer.create_quantized_param(model, param, param_name, param_device, state_dict, unexpected_keys)
932
933
934
935
936
937
938
939
            # 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)
940
            # TODO: consider removing used param_parts from state_dict before return
941
942

    return error_msgs, offload_index, state_dict_index
943
944


945
946
947
948
949
950
951
952
953
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


954
class ModuleUtilsMixin:
Julien Chaumond's avatar
Julien Chaumond committed
955
    """
956
    A few utilities for `torch.nn.Modules`, to be used as a mixin.
Julien Chaumond's avatar
Julien Chaumond committed
957
958
    """

959
960
961
962
    @staticmethod
    def _hook_rss_memory_pre_forward(module, *args, **kwargs):
        try:
            import psutil
Sylvain Gugger's avatar
Sylvain Gugger committed
963
        except ImportError:
964
965
966
967
968
969
970
971
972
973
974
            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
975
        except ImportError:
976
977
978
979
980
981
982
983
984
985
            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
986
987
988
        """
        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
989
990
        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()`.
991
992
993
994
995
996
997
        """
        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
998
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
999
        Reset the `mem_rss_diff` attribute of each module (see [`~modeling_utils.ModuleUtilsMixin.add_memory_hooks`]).
Sylvain Gugger's avatar
Sylvain Gugger committed
1000
        """
1001
1002
1003
1004
1005
        for module in self.modules():
            module.mem_rss_diff = 0
            module.mem_rss_post_forward = 0
            module.mem_rss_pre_forward = 0

1006
    @property
Sylvain Gugger's avatar
Sylvain Gugger committed
1007
    def device(self) -> torch.device:
1008
        """
1009
        `torch.device`: The device on which the module is (assuming that all the module parameters are on the same
1010
        device).
1011
        """
Lysandre Debut's avatar
Lysandre Debut committed
1012
        return get_parameter_device(self)
1013

1014
    @property
1015
    def dtype(self) -> torch.dtype:
1016
        """
1017
        `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).
1018
        """
Lysandre Debut's avatar
Lysandre Debut committed
1019
        return get_parameter_dtype(self)
1020
1021

    def invert_attention_mask(self, encoder_attention_mask: Tensor) -> Tensor:
Sylvain Gugger's avatar
Sylvain Gugger committed
1022
1023
1024
1025
        """
        Invert an attention mask (e.g., switches 0. and 1.).

        Args:
1026
            encoder_attention_mask (`torch.Tensor`): An attention mask.
Sylvain Gugger's avatar
Sylvain Gugger committed
1027
1028

        Returns:
1029
            `torch.Tensor`: The inverted attention mask.
Sylvain Gugger's avatar
Sylvain Gugger committed
1030
        """
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
        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
1041
        encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * torch.finfo(self.dtype).min
1042

1043
1044
        return encoder_extended_attention_mask

1045
    @staticmethod
1046
1047
1048
1049
1050
1051
1052
    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
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
        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

1073
    def get_extended_attention_mask(
1074
        self, attention_mask: Tensor, input_shape: Tuple[int], device: torch.device = None, dtype: torch.float = None
1075
    ) -> Tensor:
Sylvain Gugger's avatar
Sylvain Gugger committed
1076
1077
        """
        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
1078
1079

        Arguments:
1080
            attention_mask (`torch.Tensor`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1081
                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
1082
            input_shape (`Tuple[int]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1083
                The shape of the input to the model.
1084
1085

        Returns:
1086
            `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
1087
        """
Yih-Dar's avatar
Yih-Dar committed
1088
1089
1090
        if dtype is None:
            dtype = self.dtype

1091
1092
1093
1094
1095
1096
        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
                )
1097
1098
1099
1100
1101
1102
1103
1104
1105
        # 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:
1106
                extended_attention_mask = ModuleUtilsMixin.create_extended_attention_mask_for_decoder(
1107
1108
                    input_shape, attention_mask, device
                )
1109
1110
1111
1112
            else:
                extended_attention_mask = attention_mask[:, None, None, :]
        else:
            raise ValueError(
1113
                f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"
1114
1115
1116
1117
            )

        # 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
1118
        # positions we want to attend and the dtype's smallest value for masked positions.
1119
1120
        # 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
1121
1122
        extended_attention_mask = extended_attention_mask.to(dtype=dtype)  # fp16 compatibility
        extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(dtype).min
1123
1124
        return extended_attention_mask

Sylvain Gugger's avatar
Sylvain Gugger committed
1125
1126
1127
    def get_head_mask(
        self, head_mask: Optional[Tensor], num_hidden_layers: int, is_attention_chunked: bool = False
    ) -> Tensor:
1128
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
1129
1130
1131
        Prepare the head mask if needed.

        Args:
1132
            head_mask (`torch.Tensor` with shape `[num_heads]` or `[num_hidden_layers x num_heads]`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
1133
                The mask indicating if we should keep the heads or not (1.0 for keep, 0.0 for discard).
1134
            num_hidden_layers (`int`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1135
                The number of hidden layers in the model.
1136
            is_attention_chunked (`bool`, *optional*, defaults to `False`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1137
1138
                Whether or not the attentions scores are computed by chunks or not.

1139
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
1140
1141
            `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.
1142
1143
1144
        """
        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
1145
1146
            if is_attention_chunked is True:
                head_mask = head_mask.unsqueeze(-1)
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
        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()}"
1160
        head_mask = head_mask.to(dtype=self.dtype)  # switch to float if need + fp16 compatibility
1161
1162
        return head_mask

1163
1164
1165
1166
1167
    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:
1168
            only_trainable (`bool`, *optional*, defaults to `False`):
1169
1170
                Whether or not to return only the number of trainable parameters

1171
            exclude_embeddings (`bool`, *optional*, defaults to `False`):
1172
1173
1174
                Whether or not to return only the number of non-embeddings parameters

        Returns:
1175
            `int`: The number of parameters.
1176
1177
        """

1178
1179
1180
1181
        if exclude_embeddings:
            embedding_param_names = [
                f"{name}.weight" for name, module_type in self.named_modules() if isinstance(module_type, nn.Embedding)
            ]
1182
            total_parameters = [
1183
1184
1185
                parameter for name, parameter in self.named_parameters() if name not in embedding_param_names
            ]
        else:
1186
1187
1188
1189
            total_parameters = list(self.parameters())

        total_numel = []
        is_loaded_in_4bit = getattr(self, "is_loaded_in_4bit", False)
1190

1191
1192
1193
1194
1195
1196
        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"
1197
                    " make sure to install bitsandbytes with `pip install bitsandbytes`. You also need a GPU. "
1198
1199
1200
1201
1202
1203
1204
                )

        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):
1205
1206
1207
1208
1209
1210
1211
                    if hasattr(param, "element_size"):
                        num_bytes = param.element_size()
                    elif hasattr(param, "quant_storage"):
                        num_bytes = param.quant_storage.itemsize
                    else:
                        num_bytes = 1
                    total_numel.append(param.numel() * 2 * num_bytes)
1212
1213
1214
1215
                else:
                    total_numel.append(param.numel())

        return sum(total_numel)
1216
1217
1218
1219
1220
1221

    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:
1222
            inputs (`dict`): The model inputs.
1223
1224

        Returns:
1225
            `int`: The total number of tokens.
1226
        """
1227
1228
        if not hasattr(self, "warnings_issued"):
            self.warnings_issued = {}
1229
1230
        if self.main_input_name in input_dict:
            return input_dict[self.main_input_name].numel()
1231
        elif "estimate_tokens" not in self.warnings_issued:
1232
            logger.warning(
1233
1234
                "Could not estimate the number of tokens of the input, floating-point operations will not be computed"
            )
1235
1236
            self.warnings_issued["estimate_tokens"] = True
        return 0
1237
1238
1239
1240
1241
1242
1243

    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
1244
1245
        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
1246
        re-use e.g. Albert or Universal Transformers, or if doing long-range modeling with very high sequence lengths.
1247
1248

        Args:
1249
            batch_size (`int`):
1250
1251
                The batch size for the forward pass.

1252
            sequence_length (`int`):
1253
1254
                The number of tokens in each line of the batch.

1255
            exclude_embeddings (`bool`, *optional*, defaults to `True`):
1256
1257
1258
                Whether or not to count embedding and softmax operations.

        Returns:
1259
            `int`: The number of floating-point operations.
1260
1261
1262
1263
        """

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

Julien Chaumond's avatar
Julien Chaumond committed
1264

1265
class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin, PeftAdapterMixin):
1266
1267
    r"""
    Base class for all models.
1268

Sylvain Gugger's avatar
Sylvain Gugger committed
1269
1270
    [`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:
1271

1272
1273
        - resize the input embeddings,
        - prune heads in the self-attention heads.
1274

1275
    Class attributes (overridden by derived classes):
Sylvain Gugger's avatar
Sylvain Gugger committed
1276

Sylvain Gugger's avatar
Sylvain Gugger committed
1277
1278
1279
1280
        - **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:
1281

Sylvain Gugger's avatar
Sylvain Gugger committed
1282
1283
            - **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.
1284
            - **path** (`str`) -- A path to the TensorFlow checkpoint.
1285

Sylvain Gugger's avatar
Sylvain Gugger committed
1286
1287
        - **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.
1288
        - **is_parallelizable** (`bool`) -- A flag indicating whether this model supports model parallelization.
Sylvain Gugger's avatar
Sylvain Gugger committed
1289
1290
        - **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).
1291
    """
1292

1293
    config_class = None
1294
    base_model_prefix = ""
1295
    main_input_name = "input_ids"
1296
1297
    model_tags = None

1298
    _auto_class = None
1299
    _no_split_modules = None
1300
    _skip_keys_device_placement = None
1301
    _keep_in_fp32_modules = None
1302

1303
1304
    # 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.
1305
    _keys_to_ignore_on_load_missing = None
1306
1307
1308
    # 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.
1309
    _keys_to_ignore_on_load_unexpected = None
1310
1311
    # 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)
1312
    _keys_to_ignore_on_save = None
Sylvain Gugger's avatar
Sylvain Gugger committed
1313
1314
    # a list of `state_dict` keys that are potentially tied to another key in the state_dict.
    _tied_weights_keys = None
1315

1316
    is_parallelizable = False
1317
    supports_gradient_checkpointing = False
1318
    _is_stateful = False
1319

1320
1321
1322
    # Flash Attention 2 support
    _supports_flash_attn_2 = False

1323
1324
1325
    # SDPA support
    _supports_sdpa = False

1326
    # Has support for a `Cache` instance as `past_key_values`? Does it support a `StaticCache`?
1327
    _supports_cache_class = False
1328
    _supports_static_cache = False
1329

1330
1331
1332
    # Has support for a `QuantoQuantizedCache` instance as `past_key_values`
    _supports_quantized_cache = False

1333
    @property
1334
    def dummy_inputs(self) -> Dict[str, torch.Tensor]:
1335
        """
1336
        `Dict[str, torch.Tensor]`: Dummy inputs to do a forward pass in the network.
1337
        """
1338
        return {"input_ids": torch.tensor(DUMMY_INPUTS)}
1339

1340
1341
1342
1343
1344
1345
1346
    @property
    def framework(self) -> str:
        """
        :str: Identifies that this is a PyTorch model.
        """
        return "pt"

1347
    def __init__(self, config: PretrainedConfig, *inputs, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
1348
        super().__init__()
1349
1350
        if not isinstance(config, PretrainedConfig):
            raise ValueError(
1351
1352
1353
                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)`"
1354
            )
1355
        # Save config and origin of the pretrained weights if given in model
1356
1357
1358
        config = self._autoset_attn_implementation(
            config, torch_dtype=torch.get_default_dtype(), check_device_map=False
        )
1359
        self.config = config
1360

1361
        self.name_or_path = config.name_or_path
1362
        self.warnings_issued = {}
1363
        self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
1364
1365
1366
1367
        # 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)
1368
1369
1370
1371
1372
1373
1374
1375
1376

    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()

1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
    def dequantize(self):
        """
        Potentially dequantize the model in case it has been quantized by a quantization method that support
        dequantization.
        """
        hf_quantizer = getattr(self, "hf_quantizer", None)

        if hf_quantizer is None:
            raise ValueError("You need to first quantize your model in order to dequantize it")

        return hf_quantizer.dequantize(self)

1389
1390
1391
1392
1393
    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")
1394

1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
    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

1409
        model = AutoModel.from_pretrained("google-bert/bert-base-cased")
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426

        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)

1427
1428
1429
1430
1431
1432
    @classmethod
    def _from_config(cls, config, **kwargs):
        """
        All context managers that the model should be initialized under go here.

        Args:
1433
1434
            torch_dtype (`torch.dtype`, *optional*):
                Override the default `torch.dtype` and load the model under this dtype.
1435
1436
        """
        torch_dtype = kwargs.pop("torch_dtype", None)
1437
        use_flash_attention_2 = kwargs.pop("use_flash_attention_2", False)
1438
1439
1440
1441
1442
1443

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

1444
1445
1446
        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(
1447
1448
1449
1450
            config,
            use_flash_attention_2=use_flash_attention_2,
            check_device_map=False,
            torch_dtype=torch_dtype,
1451
        )
1452

1453
1454
1455
1456
1457
1458
        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
1459
            with deepspeed.zero.Init(config_dict_or_path=deepspeed_config()):
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
                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

1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
    @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)
1489
        requested_attn_implementation = None
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
        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.
1506
            requested_attn_implementation = config._attn_implementation_internal
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518

        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,
1519
                hard_check_only=False,
1520
1521
                check_device_map=check_device_map,
            )
1522
        elif requested_attn_implementation in [None, "sdpa"] and not is_torch_xla_available():
1523
            # use_flash_attention_2 takes priority over SDPA, hence SDPA treated in this elif.
1524
            config = cls._check_and_enable_sdpa(
1525
1526
                config,
                hard_check_only=False if requested_attn_implementation is None else True,
1527
            )
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537

            if (
                torch.version.hip is not None
                and config._attn_implementation == "sdpa"
                and torch.cuda.device_count() > 1
            ):
                logger.warning_once(
                    "Using the `SDPA` attention implementation on multi-gpu setup with ROCM may lead to performance issues due to the FA backend. Disabling it to use alternative backends."
                )
                torch.backends.cuda.enable_flash_sdp(False)
1538
        else:
1539
1540
1541
1542
            config._attn_implementation = "eager"

        return config

1543
1544
1545
1546
1547
1548
1549
    @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:
1550
            dtype (`torch.dtype`):
1551
1552
1553
                a floating dtype to set to.

        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
1554
1555
            `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`.
1556

1557
1558
        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.
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
        """
        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

1570
    @property
1571
1572
    def base_model(self) -> nn.Module:
        """
1573
        `torch.nn.Module`: The main body of the model.
1574
        """
1575
        return getattr(self, self.base_model_prefix, self)
thomwolf's avatar
thomwolf committed
1576

1577
1578
    @classmethod
    def can_generate(cls) -> bool:
1579
1580
1581
1582
1583
1584
        """
        Returns whether this model can generate sequences with `.generate()`.

        Returns:
            `bool`: Whether this model can generate sequences with `.generate()`.
        """
1585
1586
1587
        # 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):
1588
1589
1590
            return False
        return True

1591
1592
    @classmethod
    def _check_and_enable_flash_attn_2(
1593
1594
1595
1596
1597
1598
        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,
1599
1600
    ) -> PretrainedConfig:
        """
1601
        Checks the availability of Flash Attention 2 and compatibility with the current model.
1602

1603
        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.
1604
1605
1606
        """
        if not cls._supports_flash_attn_2:
            raise ValueError(
1607
1608
1609
                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"
1610
1611
            )

1612
        if not is_flash_attn_2_available():
1613
1614
1615
            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."

1616
1617
1618
1619
1620
            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:
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
                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}")
1634
1635
1636
1637
1638
1639
1640
1641
1642

        _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:
1643
            logger.warning_once(
1644
1645
1646
                "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]:
1647
1648
1649
1650
            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)`'
1651
1652
            )

1653
1654
1655
        # 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":
1656
            if torch.cuda.is_available():
1657
                logger.warning_once(
1658
                    "You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU"
1659
1660
1661
1662
                    " after initializing it on CPU with `model.to('cuda')`."
                )
            else:
                raise ValueError(
1663
                    "You are attempting to use Flash Attention 2.0 with a model not initialized on GPU and with no GPU available. "
1664
1665
1666
1667
                    "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 (
1668
1669
            check_device_map
            and device_map is not None
1670
1671
1672
1673
1674
1675
1676
            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."
            )
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
        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(
1691
1692
1693
                    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")`'
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
                )
            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"
1709
1710
        return config

1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
    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()

1728
    def get_input_embeddings(self) -> nn.Module:
1729
1730
1731
1732
        """
        Returns the model's input embeddings.

        Returns:
1733
            `nn.Module`: A torch module mapping vocabulary to hidden states.
thomwolf's avatar
thomwolf committed
1734
        """
1735
        base_model = getattr(self, self.base_model_prefix, self)
thomwolf's avatar
thomwolf committed
1736
1737
1738
1739
        if base_model is not self:
            return base_model.get_input_embeddings()
        else:
            raise NotImplementedError
thomwolf's avatar
thomwolf committed
1740

1741
    def set_input_embeddings(self, value: nn.Module):
1742
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
1743
        Set model's input embeddings.
1744
1745

        Args:
1746
            value (`nn.Module`): A module mapping vocabulary to hidden states.
thomwolf's avatar
thomwolf committed
1747
1748
1749
1750
1751
1752
        """
        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
1753

1754
    def get_output_embeddings(self) -> nn.Module:
1755
1756
1757
1758
        """
        Returns the model's output embeddings.

        Returns:
1759
            `nn.Module`: A torch module mapping hidden states to vocabulary.
thomwolf's avatar
thomwolf committed
1760
        """
1761
        return None  # Overwrite for models with output embeddings
thomwolf's avatar
thomwolf committed
1762

1763
1764
    def _init_weights(self, module):
        """
1765
1766
1767
1768
        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.
1769
        """
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
        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
1780

1781
    def tie_weights(self):
1782
1783
        """
        Tie the weights between the input embeddings and the output embeddings.
1784

Sylvain Gugger's avatar
Sylvain Gugger committed
1785
1786
        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
1787
        """
1788
1789
1790
1791
        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
1792

1793
        if getattr(self.config, "is_encoder_decoder", False) and getattr(self.config, "tie_encoder_decoder", False):
Weizhen's avatar
Weizhen committed
1794
1795
            if hasattr(self, self.base_model_prefix):
                self = getattr(self, self.base_model_prefix)
1796
1797
1798
1799
1800
1801
1802
            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
1803

Sylvain Gugger's avatar
Sylvain Gugger committed
1804
1805
1806
1807
        for module in self.modules():
            if hasattr(module, "_tie_weights"):
                module._tie_weights()

1808
    @staticmethod
1809
1810
1811
    def _tie_encoder_decoder_weights(
        encoder: nn.Module, decoder: nn.Module, base_model_prefix: str, base_encoder_name: str
    ):
1812
        uninitialized_encoder_weights: List[str] = []
1813
        tied_weights: List[str] = []
Weizhen's avatar
Weizhen committed
1814
1815
        if decoder.__class__ != encoder.__class__:
            logger.info(
Sylvain Gugger's avatar
Sylvain Gugger committed
1816
1817
                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
1818
            )
1819
1820
1821
1822
1823

        def tie_encoder_to_decoder_recursively(
            decoder_pointer: nn.Module,
            encoder_pointer: nn.Module,
            module_name: str,
1824
            base_encoder_name: str,
1825
1826
            uninitialized_encoder_weights: List[str],
            depth=0,
1827
1828
            total_decoder_name="",
            total_encoder_name="",
1829
1830
1831
        ):
            assert isinstance(decoder_pointer, nn.Module) and isinstance(
                encoder_pointer, nn.Module
1832
            ), f"{decoder_pointer} and {encoder_pointer} have to be of type nn.Module"
1833
1834
1835
            if hasattr(decoder_pointer, "weight"):
                assert hasattr(encoder_pointer, "weight")
                encoder_pointer.weight = decoder_pointer.weight
1836
                tied_weights.append(f"{base_encoder_name}{total_encoder_name}.weight")
1837
1838
                if hasattr(decoder_pointer, "bias"):
                    assert hasattr(encoder_pointer, "bias")
1839
                    tied_weights.append(f"{base_encoder_name}{total_encoder_name}.bias")
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
                    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}"

1850
                all_encoder_weights = {module_name + "/" + sub_name for sub_name in encoder_modules.keys()}
1851
1852
1853
1854
1855
                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
1856
1857
1858
                        if not isinstance(decoder_modules[decoder_name], type(encoder_modules[encoder_name])) and len(
                            encoder_modules
                        ) != len(decoder_modules):
1859
1860
                            # 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
1861
                            # thus skip this step and subtract one layer pos from encoder
1862
1863
1864
1865
1866
1867
                            encoder_layer_pos -= 1
                            continue
                    elif name not in encoder_modules:
                        continue
                    elif depth > 500:
                        raise ValueError(
Sylvain Gugger's avatar
Sylvain Gugger committed
1868
1869
                            "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."
1870
1871
1872
1873
1874
1875
1876
                        )
                    else:
                        decoder_name = encoder_name = name
                    tie_encoder_to_decoder_recursively(
                        decoder_modules[decoder_name],
                        encoder_modules[encoder_name],
                        module_name + "/" + name,
1877
                        base_encoder_name,
1878
1879
                        uninitialized_encoder_weights,
                        depth=depth + 1,
1880
1881
                        total_encoder_name=f"{total_encoder_name}.{encoder_name}",
                        total_decoder_name=f"{total_decoder_name}.{decoder_name}",
1882
1883
1884
1885
1886
1887
                    )
                    all_encoder_weights.remove(module_name + "/" + encoder_name)

                uninitialized_encoder_weights += list(all_encoder_weights)

        # tie weights recursively
1888
1889
1890
1891
        tie_encoder_to_decoder_recursively(
            decoder, encoder, base_model_prefix, base_encoder_name, uninitialized_encoder_weights
        )

1892
1893
1894
1895
        if len(uninitialized_encoder_weights) > 0:
            logger.warning(
                f"The following encoder weights were not tied to the decoder {uninitialized_encoder_weights}"
            )
1896
        return tied_weights
1897

1898
    def _tie_or_clone_weights(self, output_embeddings, input_embeddings):
Lysandre's avatar
Lysandre committed
1899
        """Tie or clone module weights depending of whether we are using TorchScript or not"""
thomwolf's avatar
thomwolf committed
1900
        if self.config.torchscript:
1901
            output_embeddings.weight = nn.Parameter(input_embeddings.weight.clone())
thomwolf's avatar
thomwolf committed
1902
        else:
1903
            output_embeddings.weight = input_embeddings.weight
thomwolf's avatar
thomwolf committed
1904

Sam Shleifer's avatar
Sam Shleifer committed
1905
        if getattr(output_embeddings, "bias", None) is not None:
1906
            output_embeddings.bias.data = nn.functional.pad(
1907
                output_embeddings.bias.data,
Lysandre's avatar
Lysandre committed
1908
1909
1910
1911
                (
                    0,
                    output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0],
                ),
1912
1913
                "constant",
                0,
1914
            )
1915
        if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
1916
            output_embeddings.out_features = input_embeddings.num_embeddings
1917

Marc Sun's avatar
Marc Sun committed
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
    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
        """
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
        _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
1945
1946
        return list(_no_split_modules)

1947
1948
1949
    def resize_token_embeddings(
        self, new_num_tokens: Optional[int] = None, pad_to_multiple_of: Optional[int] = None
    ) -> nn.Embedding:
1950
        """
1951
        Resizes input token embeddings matrix of the model if `new_num_tokens != config.vocab_size`.
1952

1953
        Takes care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
thomwolf's avatar
thomwolf committed
1954

1955
        Arguments:
1956
            new_num_tokens (`int`, *optional*):
1957
                The new number of tokens in the embedding matrix. Increasing the size will add newly initialized
Sylvain Gugger's avatar
Sylvain Gugger committed
1958
1959
                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.
1960
            pad_to_multiple_of (`int`, *optional*):
Arthur's avatar
Arthur committed
1961
1962
                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`.
1963
1964
1965
1966
1967

                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
1968
1969

        Return:
1970
            `torch.nn.Embedding`: Pointer to the input tokens Embeddings Module of the model.
thomwolf's avatar
thomwolf committed
1971
        """
1972
        model_embeds = self._resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
Arthur's avatar
Arthur committed
1973
        if new_num_tokens is None and pad_to_multiple_of is None:
thomwolf's avatar
thomwolf committed
1974
            return model_embeds
thomwolf's avatar
thomwolf committed
1975
1976

        # Update base model and current model config
1977
1978
        if hasattr(self.config, "text_config"):
            self.config.text_config.vocab_size = model_embeds.weight.shape[0]
1979
1980
        else:
            self.config.vocab_size = model_embeds.weight.shape[0]
Arthur's avatar
Arthur committed
1981
        self.vocab_size = model_embeds.weight.shape[0]
thomwolf's avatar
thomwolf committed
1982
1983

        # Tie weights again if needed
1984
        self.tie_weights()
thomwolf's avatar
thomwolf committed
1985

thomwolf's avatar
thomwolf committed
1986
1987
        return model_embeds

1988
    def _resize_token_embeddings(self, new_num_tokens, pad_to_multiple_of=None):
thomwolf's avatar
thomwolf committed
1989
        old_embeddings = self.get_input_embeddings()
1990
        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens, pad_to_multiple_of)
1991
1992
1993
        if hasattr(old_embeddings, "_hf_hook"):
            hook = old_embeddings._hf_hook
            add_hook_to_module(new_embeddings, hook)
1994
1995
        old_embeddings_requires_grad = old_embeddings.weight.requires_grad
        new_embeddings.requires_grad_(old_embeddings_requires_grad)
thomwolf's avatar
thomwolf committed
1996
        self.set_input_embeddings(new_embeddings)
1997
        is_quantized = hasattr(self, "hf_quantizer") and self.hf_quantizer is not None
1998

1999
2000
        # Update new_num_tokens with the actual size of new_embeddings
        if pad_to_multiple_of is not None:
2001
            if is_deepspeed_zero3_enabled() and not is_quantized:
2002
2003
2004
2005
2006
2007
2008
                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]

2009
2010
2011
        # 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()
2012
2013
2014
2015
            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)
2016
2017
2018
            if hasattr(old_lm_head, "_hf_hook"):
                hook = old_lm_head._hf_hook
                add_hook_to_module(new_lm_head, hook)
2019
2020
            old_lm_head_requires_grad = old_lm_head.weight.requires_grad
            new_lm_head.requires_grad_(old_lm_head_requires_grad)
2021
2022
            self.set_output_embeddings(new_lm_head)

thomwolf's avatar
thomwolf committed
2023
        return self.get_input_embeddings()
2024

2025
    def _get_resized_embeddings(
2026
2027
2028
2029
        self,
        old_embeddings: nn.Embedding,
        new_num_tokens: Optional[int] = None,
        pad_to_multiple_of: Optional[int] = None,
2030
    ) -> nn.Embedding:
2031
2032
2033
        """
        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
2034
2035

        Args:
2036
            old_embeddings (`torch.nn.Embedding`):
2037
                Old embeddings to be resized.
2038
            new_num_tokens (`int`, *optional*):
2039
                New number of tokens in the embedding matrix.
2040
2041

                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
2042
                vectors from the end. If not provided or `None`, just returns a pointer to the input tokens
2043
                `torch.nn.Embedding` module of the model without doing anything.
2044
            pad_to_multiple_of (`int`, *optional*):
Arthur's avatar
Arthur committed
2045
2046
                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`.
2047
2048
2049
2050
2051
2052

                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

2053
2054

        Return:
2055
2056
            `torch.nn.Embedding`: Pointer to the resized Embedding Module or the old Embedding Module if
            `new_num_tokens` is `None`
2057
        """
2058
2059
2060
2061
2062
2063
2064
2065

        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]
2066
            new_num_tokens = ((new_num_tokens + pad_to_multiple_of - 1) // pad_to_multiple_of) * pad_to_multiple_of
2067
        else:
2068
            logger.info(
2069
                "You are resizing the embedding layer without providing a `pad_to_multiple_of` parameter. This means that the new embedding"
2070
                f" dimension will be {new_num_tokens}. This might induce some performance reduction as *Tensor Cores* will not be available."
2071
                " For more details about this, or help on choosing the correct value for resizing, refer to this guide:"
2072
2073
2074
                " https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html#requirements-tc"
            )

2075
2076
2077
        if new_num_tokens is None:
            return old_embeddings

2078
2079
        is_quantized = hasattr(self, "hf_quantizer") and self.hf_quantizer is not None
        if is_deepspeed_zero3_enabled() and not is_quantized:
2080
2081
2082
2083
2084
2085
2086
            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()

2087
        if old_num_tokens == new_num_tokens and not is_deepspeed_zero3_enabled():
2088
2089
            return old_embeddings

2090
2091
        if not isinstance(old_embeddings, nn.Embedding):
            raise TypeError(
Sylvain Gugger's avatar
Sylvain Gugger committed
2092
2093
2094
                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}."
2095
2096
            )

2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
        # 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

2115
2116
        # numbers of tokens to copy
        n = min(old_num_tokens, new_num_tokens)
2117

2118
        if is_deepspeed_zero3_enabled() and not is_quantized:
2119
2120
            import deepspeed

2121
2122
2123
            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, :]
2124
2125
        else:
            new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
2126
2127
2128

        return new_embeddings

2129
    def _get_resized_lm_head(
2130
2131
        self, old_lm_head: nn.Linear, new_num_tokens: Optional[int] = None, transposed: Optional[bool] = False
    ) -> nn.Linear:
2132
2133
2134
2135
2136
        """
        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:
2137
            old_lm_head (`torch.nn.Linear`):
2138
                Old lm head liner layer to be resized.
2139
            new_num_tokens (`int`, *optional*):
2140
2141
2142
                New number of tokens in the linear matrix.

                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
2143
                vectors from the end. If not provided or `None`, just returns a pointer to the input tokens
2144
2145
2146
                `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`.
2147
2148

        Return:
Sylvain Gugger's avatar
Sylvain Gugger committed
2149
2150
            `torch.nn.Linear`: Pointer to the resized Linear Module or the old Linear Module if `new_num_tokens` is
            `None`
2151
2152
2153
2154
        """
        if new_num_tokens is None:
            return old_lm_head

2155
2156
        is_quantized = hasattr(self, "hf_quantizer") and self.hf_quantizer is not None
        if is_deepspeed_zero3_enabled() and not is_quantized:
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
            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()
            )
2167

2168
        if old_num_tokens == new_num_tokens and not is_deepspeed_zero3_enabled():
2169
2170
2171
2172
            return old_lm_head

        if not isinstance(old_lm_head, nn.Linear):
            raise TypeError(
Sylvain Gugger's avatar
Sylvain Gugger committed
2173
2174
2175
                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}."
2176
2177
2178
2179
2180
2181
            )

        # 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

2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
        # 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)

2196
2197
        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)

2198
        if is_deepspeed_zero3_enabled() and not is_quantized:
2199
2200
            import deepspeed

2201
2202
            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):
2203
2204
2205
                self._copy_lm_head_original_to_resized(
                    new_lm_head, old_lm_head, num_tokens_to_copy, transposed, has_new_lm_head_bias
                )
2206
        else:
2207
2208
            self._copy_lm_head_original_to_resized(
                new_lm_head, old_lm_head, num_tokens_to_copy, transposed, has_new_lm_head_bias
2209
            )
2210
2211
2212

        return new_lm_head

2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
    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]

2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
    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`"
        )

2238
    def init_weights(self):
2239
        """
2240
2241
        If needed prunes and maybe initializes weights. If using a custom `PreTrainedModel`, you need to implement any
        initialization logic in `_init_weights`.
2242
        """
2243
2244
2245
2246
        # Prune heads if needed
        if self.config.pruned_heads:
            self.prune_heads(self.config.pruned_heads)

2247
2248
        if _init_weights:
            # Initialize weights
2249
            self.apply(self._initialize_weights)
2250
2251
2252
2253

            # Tie weights should be skipped when not initializing all weights
            # since from_pretrained(...) calls tie weights anyways
            self.tie_weights()
2254

2255
2256
2257
    def prune_heads(self, heads_to_prune: Dict[int, List[int]]):
        """
        Prunes heads of the base model.
2258

2259
        Arguments:
2260
            heads_to_prune (`Dict[int, List[int]]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2261
2262
2263
                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
2264
        """
2265
        # save new sets of pruned heads as union of previously stored pruned heads and newly pruned heads
2266
        for layer, heads in heads_to_prune.items():
2267
2268
2269
            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

2270
        self.base_model._prune_heads(heads_to_prune)
thomwolf's avatar
thomwolf committed
2271

2272
    def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):
2273
2274
2275
2276
2277
        """
        Activates gradient checkpointing for the current model.

        Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint
        activations".
2278
2279
2280
2281
2282
2283
2284

        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.
2285
2286
2287
        """
        if not self.supports_gradient_checkpointing:
            raise ValueError(f"{self.__class__.__name__} does not support gradient checkpointing.")
2288
2289

        if gradient_checkpointing_kwargs is None:
2290
            gradient_checkpointing_kwargs = {"use_reentrant": True}
2291

2292
        gradient_checkpointing_func = functools.partial(checkpoint, **gradient_checkpointing_kwargs)
2293

2294
        # For old GC format (transformers < 4.35.0) for models that live on the Hub
Stas Bekman's avatar
Stas Bekman committed
2295
        # we will fall back to the overwritten `_set_gradient_checkpointing` method
2296
2297
2298
2299
2300
2301
        _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))
2302
            logger.warning(
2303
2304
2305
                "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."
            )
2306

2307
2308
2309
2310
2311
2312
2313
        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()

2314
    def _set_gradient_checkpointing(self, enable: bool = True, gradient_checkpointing_func: Callable = checkpoint):
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
        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."
            )

2336
    def gradient_checkpointing_disable(self):
2337
2338
2339
2340
2341
2342
2343
        """
        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:
2344
2345
2346
2347
2348
2349
            # 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:
2350
                logger.warning(
2351
2352
2353
2354
                    "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))
2355

2356
2357
2358
        if getattr(self, "_hf_peft_config_loaded", False):
            self.disable_input_require_grads()

2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
    @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())

2369
2370
2371
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
2372
        is_main_process: bool = True,
2373
2374
        state_dict: Optional[dict] = None,
        save_function: Callable = torch.save,
Sylvain Gugger's avatar
Sylvain Gugger committed
2375
        push_to_hub: bool = False,
2376
        max_shard_size: Union[int, str] = "5GB",
2377
        safe_serialization: bool = True,
2378
        variant: Optional[str] = None,
2379
        token: Optional[Union[str, bool]] = None,
2380
        save_peft_format: bool = True,
Sylvain Gugger's avatar
Sylvain Gugger committed
2381
        **kwargs,
2382
    ):
2383
2384
        """
        Save a model and its configuration file to a directory, so that it can be re-loaded using the
2385
        [`~PreTrainedModel.from_pretrained`] class method.
2386

2387
        Arguments:
2388
            save_directory (`str` or `os.PathLike`):
2389
                Directory to which to save. Will be created if it doesn't exist.
2390
2391
2392
2393
            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.
2394
            state_dict (nested dictionary of `torch.Tensor`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2395
2396
2397
                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).
2398
            save_function (`Callable`):
2399
                The function to use to save the state dictionary. Useful on distributed training like TPUs when one
2400
2401
                need to replace `torch.save` by another method.
            push_to_hub (`bool`, *optional*, defaults to `False`):
2402
2403
2404
                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).
2405
            max_shard_size (`int` or `str`, *optional*, defaults to `"5GB"`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2406
2407
                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"`).
2408
2409
                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
2410
2411
2412
2413
2414
2415
2416
2417

                <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>

2418
            safe_serialization (`bool`, *optional*, defaults to `True`):
2419
                Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
2420
2421
            variant (`str`, *optional*):
                If specified, weights are saved in the format pytorch_model.<variant>.bin.
2422
2423
2424
            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`).
2425
2426
2427
2428
            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`.
2429
            kwargs (`Dict[str, Any]`, *optional*):
2430
                Additional key word arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
2431
        """
2432
        use_auth_token = kwargs.pop("use_auth_token", None)
2433
        ignore_metadata_errors = kwargs.pop("ignore_metadata_errors", False)
2434
2435
2436

        if use_auth_token is not None:
            warnings.warn(
2437
2438
                "The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
                FutureWarning,
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
            )
            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
2449
2450
        _hf_peft_config_loaded = getattr(self, "_hf_peft_config_loaded", False)

2451
2452
2453
2454
        hf_quantizer = getattr(self, "hf_quantizer", None)
        quantization_serializable = (
            hf_quantizer is not None and isinstance(hf_quantizer, HfQuantizer) and hf_quantizer.is_serializable
        )
2455

2456
2457
2458
2459
        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."
2460
2461
            )

2462
2463
2464
2465
2466
        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")
2467
2468
        if safe_serialization and not is_safetensors_available():
            raise ImportError("`safe_serialization` requires the `safetensors library: `pip install safetensors`.")
2469

2470
        if os.path.isfile(save_directory):
2471
            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
2472
            return
2473

2474
2475
        os.makedirs(save_directory, exist_ok=True)

2476
2477
        if push_to_hub:
            commit_message = kwargs.pop("commit_message", None)
2478
            repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
2479
            repo_id = self._create_repo(repo_id, **kwargs)
2480
            files_timestamps = self._get_files_timestamps(save_directory)
2481

Julien Chaumond's avatar
Julien Chaumond committed
2482
        # Only save the model itself if we are using distributed training
2483
        model_to_save = unwrap_model(self)
2484

2485
2486
2487
2488
2489
        # 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
2490
2491
2492
        # Attach architecture to the config
        model_to_save.config.architectures = [model_to_save.__class__.__name__]

2493
2494
2495
2496
2497
        # 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)

2498
        # Save the config
2499
        if is_main_process:
2500
2501
            if not _hf_peft_config_loaded:
                model_to_save.config.save_pretrained(save_directory)
2502
            if self.can_generate():
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
                # 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."
                        )
2521
                model_to_save.generation_config.save_pretrained(save_directory)
2522

2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
            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

2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
                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]
2548
2549
                current_peft_config.save_pretrained(save_directory)

2550
2551
2552
        # for offloaded modules
        module_map = {}

2553
2554
        # Save the model
        if state_dict is None:
2555
2556
2557
2558
2559
            # if any model parameters are offloaded, make module map
            if (
                hasattr(self, "hf_device_map")
                and len(set(self.hf_device_map.values())) > 1
                and ("cpu" in self.hf_device_map.values() or "disk" in self.hf_device_map.values())
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
            ):
                warnings.warn(
                    "Attempting to save a model with offloaded modules. Ensure that unallocated cpu memory exceeds the `shard_size` (5GB default)"
                )
                for name, module in model_to_save.named_modules():
                    if name == "":
                        continue
                    module_state_dict = module.state_dict()

                    for key in module_state_dict:
                        module_map[name + f".{key}"] = module
2571
            state_dict = model_to_save.state_dict()
2572

2573
2574
2575
2576
2577
        # 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)

2578
        # Handle the case where some state_dict keys shouldn't be saved
2579
        if self._keys_to_ignore_on_save is not None:
2580
            for ignore_key in self._keys_to_ignore_on_save:
2581
2582
                if ignore_key in state_dict.keys():
                    del state_dict[ignore_key]
2583
2584
2585
2586
2587
        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():
2588
2589
2590
2591
2592
2593
2594
                # 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)
2595

2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
            # These are all the pointers of shared tensors
            if hasattr(self, "hf_device_map"):
                # if the model has offloaded parameters, we must check using find_tied_parameters()
                tied_params = find_tied_parameters(self)
                if tied_params:
                    tied_names = tied_params[0]
                    shared_ptrs = {
                        ptr: names for ptr, names in ptrs.items() if any(name in tied_names for name in names)
                    }
                else:
                    shared_ptrs = {}
            else:
                shared_ptrs = {ptr: names for ptr, names in ptrs.items() if len(names) > 1}

2610
2611
            # Recursively descend to find tied weight keys
            _tied_weights_keys = _get_tied_weight_keys(self)
2612
2613
            error_names = []
            to_delete_names = set()
2614
2615
2616
            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.
2617
                if _tied_weights_keys is not None:
2618
2619
                    found = 0
                    for name in sorted(names):
2620
                        matches_pattern = any(re.search(pat, name) for pat in _tied_weights_keys)
2621
                        if matches_pattern and name in state_dict:
2622
2623
                            found += 1
                            if found < len(names):
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
                                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.",
2653
                )
2654

Sylvain Gugger's avatar
Sylvain Gugger committed
2655
        # Shard the model if it is too big.
2656
2657
2658
2659
2660
        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
2661

2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
        state_dict_split = split_torch_state_dict_into_shards(
            state_dict, filename_pattern=filename_pattern, max_shard_size=max_shard_size
        )
        # Save index if sharded
        index = None
        if state_dict_split.is_sharded:
            index = {
                "metadata": state_dict_split.metadata,
                "weight_map": state_dict_split.tensor_to_filename,
            }
Sylvain Gugger's avatar
Sylvain Gugger committed
2673
2674
2675
2676

        # Clean the folder from a previous save
        for filename in os.listdir(save_directory):
            full_filename = os.path.join(save_directory, filename)
2677
2678
            # 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.
2679
            weights_no_suffix = weights_name.replace(".bin", "").replace(".safetensors", "")
2680
2681
2682

            # 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", "")
2683
            reg = re.compile(r"(.*?)-\d{5}-of-\d{5}")
2684

2685
            if (
2686
                filename.startswith(weights_no_suffix)
2687
                and os.path.isfile(full_filename)
2688
                and filename not in state_dict_split.filename_to_tensors.keys()
2689
                and is_main_process
2690
                and reg.fullmatch(filename_no_suffix) is not None
2691
            ):
Sylvain Gugger's avatar
Sylvain Gugger committed
2692
2693
                os.remove(full_filename)
        # Save the model
2694
2695
2696
2697
        filename_to_tensors = state_dict_split.filename_to_tensors.items()
        if module_map:
            filename_to_tensors = logging.tqdm(filename_to_tensors, desc="Saving checkpoint shards")
        for shard_file, tensors in filename_to_tensors:
2698
            shard = {tensor: state_dict[tensor] for tensor in tensors}
2699
2700
2701
2702
2703
2704
2705
2706
            # remake shard with onloaded parameters if necessary
            if module_map:
                if accelerate_version < version.parse("0.31"):
                    raise ImportError(
                        f"You need accelerate version to be greater or equal than 0.31 to save models with offloaded parameters. Detected version {accelerate_version}. "
                        f"Please upgrade accelerate with `pip install -U accelerate`"
                    )
                # init state_dict for this shard
2707
                shard_state_dict = {name: "" for name in shard}
2708
2709
2710
                for module_name in shard:
                    module = module_map[module_name]
                    # update state dict with onloaded parameters
2711
                    shard_state_dict = get_state_dict_from_offload(module, module_name, shard_state_dict)
2712
2713

                # assign shard to be the completed state dict
2714
2715
                shard = shard_state_dict
                del shard_state_dict
2716
2717
                gc.collect()

2718
2719
2720
2721
2722
2723
            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
2724
2725

        if index is None:
2726
            path_to_weights = os.path.join(save_directory, weights_name)
2727
            logger.info(f"Model weights saved in {path_to_weights}")
Sylvain Gugger's avatar
Sylvain Gugger committed
2728
        else:
2729
            save_index_file = SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
2730
            save_index_file = os.path.join(save_directory, _add_variant(save_index_file, variant))
Sylvain Gugger's avatar
Sylvain Gugger committed
2731
2732
2733
2734
2735
2736
            # 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 "
2737
                f"split in {len(state_dict_split.filename_to_tensors)} checkpoint shards. You can find where each parameters has been saved in the "
Sylvain Gugger's avatar
Sylvain Gugger committed
2738
2739
                f"index located at {save_index_file}."
            )
2740

Sylvain Gugger's avatar
Sylvain Gugger committed
2741
        if push_to_hub:
2742
2743
2744
2745
2746
2747
2748
2749
            # 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"))

2750
            self._upload_modified_files(
2751
2752
2753
2754
                save_directory,
                repo_id,
                files_timestamps,
                commit_message=commit_message,
2755
                token=token,
2756
            )
Sylvain Gugger's avatar
Sylvain Gugger committed
2757

2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
    @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)

2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
    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

2792
    @wraps(torch.nn.Module.cuda)
2793
    def cuda(self, *args, **kwargs):
2794
2795
        if getattr(self, "quantization_method", None) == QuantizationMethod.HQQ:
            raise ValueError("`.cuda` is not supported for HQQ-quantized models.")
2796
        # Checks if the model has been loaded in 8-bit
Marc Sun's avatar
Marc Sun committed
2797
        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
2798
2799
2800
2801
2802
2803
2804
            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)

2805
    @wraps(torch.nn.Module.to)
2806
    def to(self, *args, **kwargs):
2807
2808
        if getattr(self, "quantization_method", None) == QuantizationMethod.HQQ:
            raise ValueError("`.to` is not supported for HQQ-quantized models.")
2809
        # Checks if the model has been loaded in 8-bit
Marc Sun's avatar
Marc Sun committed
2810
        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
2811
            raise ValueError(
Marc Sun's avatar
Marc Sun committed
2812
                "`.to` is not supported for `4-bit` or `8-bit` bitsandbytes models. Please use the model as it is, since the"
2813
2814
                " model has already been set to the correct devices and casted to the correct `dtype`."
            )
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
        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)
2834
2835

    def half(self, *args):
Marc Sun's avatar
Marc Sun committed
2836
        # Checks if the model is quantized
2837
        if getattr(self, "is_quantized", False):
2838
            raise ValueError(
Marc Sun's avatar
Marc Sun committed
2839
                "`.half()` is not supported for quantized model. Please use the model as it is, since the"
2840
2841
2842
2843
2844
2845
                " 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
2846
        # Checks if the model is quantized
2847
        if getattr(self, "is_quantized", False):
2848
            raise ValueError(
Marc Sun's avatar
Marc Sun committed
2849
                "`.float()` is not supported for quantized model. Please use the model as it is, since the"
2850
2851
2852
2853
2854
                " model has already been casted to the correct `dtype`."
            )
        else:
            return super().float(*args)

2855
    @classmethod
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
    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,
2869
    ) -> "PreTrainedModel":
2870
2871
        r"""
        Instantiate a pretrained pytorch model from a pre-trained model configuration.
2872

Sylvain Gugger's avatar
Sylvain Gugger committed
2873
2874
        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()`.
2875

2876
        The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
2877
2878
        pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
        task.
2879

2880
        The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
2881
        weights are discarded.
2882

2883
2884
2885
2886
        If model weights are the same precision as the base model (and is a supported model), weights will be lazily loaded
        in using the `meta` device and brought into memory once an input is passed through that layer regardless of
        `low_cpu_mem_usage`.

2887
        Parameters:
2888
            pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
2889
2890
                Can be either:

2891
2892
2893
2894
                    - 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
2895
2896
2897
                      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.
2898
                    - 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
2899
2900
                      `./flax_model/` containing `flax_model.msgpack`). In this case, `from_flax` should be set to
                      `True`.
2901
2902
2903
2904
2905
                    - `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*):
2906
2907
                Can be either:

2908
2909
                    - an instance of a class derived from [`PretrainedConfig`],
                    - a string or path valid as input to [`~PretrainedConfig.from_pretrained`].
2910

2911
                Configuration for the model to use instead of an automatically loaded configuration. Configuration can
2912
2913
                be automatically loaded when:

2914
                    - The model is a model provided by the library (loaded with the *model id* string of a pretrained
2915
                      model).
Sylvain Gugger's avatar
Sylvain Gugger committed
2916
2917
                    - The model was saved using [`~PreTrainedModel.save_pretrained`] and is reloaded by supplying the
                      save directory.
2918
2919
2920
                    - 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*):
2921
2922
2923
                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
2924
                weights. In this case though, you should check if using [`~PreTrainedModel.save_pretrained`] and
2925
2926
                [`~PreTrainedModel.from_pretrained`] is not a simpler option.
            cache_dir (`Union[str, os.PathLike]`, *optional*):
2927
2928
                Path to a directory in which a downloaded pretrained model configuration should be cached if the
                standard cache should not be used.
2929
            from_tf (`bool`, *optional*, defaults to `False`):
2930
                Load the model weights from a TensorFlow checkpoint save file (see docstring of
2931
2932
                `pretrained_model_name_or_path` argument).
            from_flax (`bool`, *optional*, defaults to `False`):
2933
                Load the model weights from a Flax checkpoint save file (see docstring of
2934
2935
                `pretrained_model_name_or_path` argument).
            ignore_mismatched_sizes (`bool`, *optional*, defaults to `False`):
2936
2937
2938
                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).
2939
            force_download (`bool`, *optional*, defaults to `False`):
2940
2941
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
2942
2943
2944
            resume_download:
                Deprecated and ignored. All downloads are now resumed by default when possible.
                Will be removed in v5 of Transformers.
2945
            proxies (`Dict[str, str]`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
2946
2947
                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.
2948
            output_loading_info(`bool`, *optional*, defaults to `False`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2949
                Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
2950
            local_files_only(`bool`, *optional*, defaults to `False`):
Stas Bekman's avatar
Stas Bekman committed
2951
                Whether or not to only look at local files (i.e., do not try to download the model).
2952
            token (`str` or `bool`, *optional*):
2953
2954
                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`).
2955
            revision (`str`, *optional*, defaults to `"main"`):
Julien Chaumond's avatar
Julien Chaumond committed
2956
                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
2957
                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
Julien Chaumond's avatar
Julien Chaumond committed
2958
                identifier allowed by git.
2959
2960
2961
2962
2963
2964
2965

                <Tip>

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

                </Tip>

2966
            mirror (`str`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
2967
2968
2969
                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.
2970
            _fast_init(`bool`, *optional*, defaults to `True`):
2971
2972
                Whether or not to disable fast initialization.

2973
2974
                <Tip warning={true}>

Sylvain Gugger's avatar
Sylvain Gugger committed
2975
2976
2977
                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.
2978

2979
                </Tip>
2980
2981
            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.
2982

2983
2984
2985
            > Parameters for big model inference

            low_cpu_mem_usage(`bool`, *optional*):
2986
                Tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
2987
                Generally should be combined with a `device_map` (such as `"auto"`) for best results.
2988
                This is an experimental feature and a subject to change at any moment.
2989
2990
2991
2992
2993
                </Tip>
                    If the model weights are in the same precision as the model loaded in, `low_cpu_mem_usage` (without
                    `device_map`) is redundant and will not provide any benefit in regards to CPU memory usage. However,
                    this should still be enabled if you are passing in a `device_map`.
                </Tip>
2994
            torch_dtype (`str` or `torch.dtype`, *optional*):
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
                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.

3008
3009
                3. A string that is a valid `torch.dtype`. E.g. "float32" loads the model in `torch.float32`, "float16" loads in `torch.float16` etc.

3010
3011
3012
3013
3014
3015
3016
3017
                <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>

3018
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]` or `int` or `torch.device`, *optional*):
3019
3020
                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
3021
3022
3023
                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.
3024

3025
3026
                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
3027
                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
3028
3029
3030
            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.
3031
3032
            offload_folder (`str` or `os.PathLike`, *optional*):
                If the `device_map` contains any value `"disk"`, the folder where we will offload weights.
3033
            offload_state_dict (`bool`, *optional*):
3034
                If `True`, will temporarily offload the CPU state dict to the hard drive to avoid getting out of CPU
3035
3036
                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.
3037
3038
            offload_buffers (`bool`, *optional*):
                Whether or not to offload the buffers with the model parameters.
Marc Sun's avatar
Marc Sun committed
3039
3040
            quantization_config (`Union[QuantizationConfigMixin,Dict]`, *optional*):
                A dictionary of configuration parameters or a QuantizationConfigMixin object for quantization (e.g
3041
3042
3043
3044
                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.
3045
3046
3047
            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.
3048
3049
3050
            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`.
3051
3052
3053
            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`.
3054

3055
            kwargs (remaining dictionary of keyword arguments, *optional*):
3056
                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
3057
                `output_attentions=True`). Behaves differently depending on whether a `config` is provided or
3058
3059
                automatically loaded:

3060
3061
                    - 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
3062
                      already been done)
3063
                    - If a configuration is not provided, `kwargs` will be first passed to the configuration class
Sylvain Gugger's avatar
Sylvain Gugger committed
3064
3065
3066
3067
                      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.
3068
3069
3070

        <Tip>

Sylvain Gugger's avatar
Sylvain Gugger committed
3071
3072
        Activate the special ["offline-mode"](https://huggingface.co/transformers/installation.html#offline-mode) to
        use this method in a firewalled environment.
3073
3074
3075
3076
3077
3078
3079

        </Tip>

        Examples:

        ```python
        >>> from transformers import BertConfig, BertModel
Sylvain Gugger's avatar
Sylvain Gugger committed
3080

3081
        >>> # Download model and configuration from huggingface.co and cache.
3082
        >>> model = BertModel.from_pretrained("google-bert/bert-base-uncased")
3083
        >>> # Model was saved using *save_pretrained('./test/saved_model/')* (for example purposes, not runnable).
Sylvain Gugger's avatar
Sylvain Gugger committed
3084
        >>> model = BertModel.from_pretrained("./test/saved_model/")
3085
        >>> # Update configuration during loading.
3086
        >>> model = BertModel.from_pretrained("google-bert/bert-base-uncased", output_attentions=True)
3087
3088
        >>> 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
3089
3090
        >>> 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)
3091
        >>> # Loading from a Flax checkpoint file instead of a PyTorch model (slower)
3092
        >>> model = BertModel.from_pretrained("google-bert/bert-base-uncased", from_flax=True)
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
        ```

        * `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

        """
3111
3112
        state_dict = kwargs.pop("state_dict", None)
        from_tf = kwargs.pop("from_tf", False)
3113
        from_flax = kwargs.pop("from_flax", False)
3114
        resume_download = kwargs.pop("resume_download", None)
3115
3116
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
3117
        use_auth_token = kwargs.pop("use_auth_token", None)
3118
        trust_remote_code = kwargs.pop("trust_remote_code", None)
Sylvain Gugger's avatar
Sylvain Gugger committed
3119
        _ = kwargs.pop("mirror", None)
3120
3121
        from_pipeline = kwargs.pop("_from_pipeline", None)
        from_auto_class = kwargs.pop("_from_auto", False)
3122
        _fast_init = kwargs.pop("_fast_init", True)
3123
        torch_dtype = kwargs.pop("torch_dtype", None)
3124
3125
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", None)
        device_map = kwargs.pop("device_map", None)
3126
        max_memory = kwargs.pop("max_memory", None)
3127
        offload_folder = kwargs.pop("offload_folder", None)
3128
        offload_state_dict = kwargs.pop("offload_state_dict", False)
3129
        offload_buffers = kwargs.pop("offload_buffers", False)
3130
        load_in_8bit = kwargs.pop("load_in_8bit", False)
3131
        load_in_4bit = kwargs.pop("load_in_4bit", False)
3132
        quantization_config = kwargs.pop("quantization_config", None)
3133
        subfolder = kwargs.pop("subfolder", "")
3134
        commit_hash = kwargs.pop("_commit_hash", None)
3135
        variant = kwargs.pop("variant", None)
3136
        adapter_kwargs = kwargs.pop("adapter_kwargs", {})
3137
        adapter_name = kwargs.pop("adapter_name", "default")
3138
        use_flash_attention_2 = kwargs.pop("use_flash_attention_2", False)
3139

3140
3141
3142
3143
        gguf_file = kwargs.pop("gguf_file", None)
        # Cache path to the GGUF file
        gguf_path = None

3144
3145
3146
        if is_fsdp_enabled():
            low_cpu_mem_usage = True

3147
3148
        if use_auth_token is not None:
            warnings.warn(
3149
3150
                "The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
                FutureWarning,
3151
3152
3153
3154
3155
3156
3157
            )
            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

3158
        if token is not None and adapter_kwargs is not None and "token" not in adapter_kwargs:
3159
3160
            adapter_kwargs["token"] = token

3161
3162
        if use_safetensors is None and not is_safetensors_available():
            use_safetensors = False
3163
3164
3165
3166
3167
        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."
            )
3168

3169
3170
3171
        if gguf_file is not None and not is_accelerate_available():
            raise ValueError("accelerate is required when loading a GGUF file `pip install accelerate`.")

3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
        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,
3186
                    _raise_exceptions_for_gated_repo=False,
3187
3188
3189
3190
3191
3192
3193
                    _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)

3194
        if is_peft_available():
3195
3196
            _adapter_model_path = adapter_kwargs.pop("_adapter_model_path", None)

3197
3198
3199
3200
3201
3202
3203
3204
3205
            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,
3206
                    **adapter_kwargs,
3207
3208
                )
            if _adapter_model_path is not None and os.path.isfile(_adapter_model_path):
3209
                with open(_adapter_model_path, "r", encoding="utf-8") as f:
3210
                    _adapter_model_path = pretrained_model_name_or_path
3211
                    pretrained_model_name_or_path = json.load(f)["base_model_name_or_path"]
3212
3213
        else:
            _adapter_model_path = None
3214

3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
        # 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}

3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
        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`"
                )
3249

3250
3251
3252
        # 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:
3253
                raise ValueError(
3254
                    "You can't pass `load_in_4bit`or `load_in_8bit` as a kwarg when passing "
3255
3256
3257
                    "`quantization_config` argument at the same time."
                )

3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
            # 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."
            )
3268

3269
        from_pt = not (from_tf | from_flax)
3270

3271
3272
3273
        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
3274

3275
3276
3277
3278
        if is_offline_mode() and not local_files_only:
            logger.info("Offline mode: forcing local_files_only=True")
            local_files_only = True

3279
3280
3281
        # 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
3282
            config, model_kwargs = cls.config_class.from_pretrained(
3283
3284
3285
                config_path,
                cache_dir=cache_dir,
                return_unused_kwargs=True,
3286
                force_download=force_download,
3287
                resume_download=resume_download,
3288
                proxies=proxies,
3289
                local_files_only=local_files_only,
3290
                token=token,
Julien Chaumond's avatar
Julien Chaumond committed
3291
                revision=revision,
3292
                subfolder=subfolder,
3293
3294
                _from_auto=from_auto_class,
                _from_pipeline=from_pipeline,
3295
                **kwargs,
3296
3297
            )
        else:
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
            # 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)
3308
            if kwarg_attn_imp is not None:
3309
                config._attn_implementation = kwarg_attn_imp
3310

3311
            model_kwargs = kwargs
3312

3313
3314
3315
3316
3317
        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
3318
3319
3320
                )
            else:
                config.quantization_config = quantization_config
3321
3322
3323
            hf_quantizer = AutoHfQuantizer.from_config(config.quantization_config, pre_quantized=pre_quantized)
        else:
            hf_quantizer = None
3324

3325
3326
3327
3328
3329
3330
        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)
3331

3332
3333
3334
            # In order to ensure popular quantization methods are supported. Can be disable with `disable_telemetry`
            user_agent["quant"] = hf_quantizer.quantization_config.quant_method.value

3335
3336
3337
            # Force-set to `True` for more mem efficiency
            if low_cpu_mem_usage is None:
                low_cpu_mem_usage = True
3338
                logger.warning("`low_cpu_mem_usage` was None, now set to True since model is quantized.")
3339
        is_quantized = hf_quantizer is not None
3340

Sylvain Gugger's avatar
Sylvain Gugger committed
3341
3342
3343
3344
        # 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
3345
        # Load model
Yih-Dar's avatar
Yih-Dar committed
3346
3347
        loading_info = None

3348
3349
3350
3351
        # Keep in fp32 modules
        keep_in_fp32_modules = None
        use_keep_in_fp32_modules = False

3352
3353
3354
3355
3356
3357
        if gguf_file is not None and hf_quantizer is not None:
            raise ValueError(
                "You cannot combine Quantization and loading a model from a GGUF file, try again by making sure you did not passed a `quantization_config` or that you did not load a quantized model from the Hub."
            )

        if pretrained_model_name_or_path is not None and gguf_file is None:
3358
            pretrained_model_name_or_path = str(pretrained_model_name_or_path)
3359
3360
            is_local = os.path.isdir(pretrained_model_name_or_path)
            if is_local:
3361
3362
3363
                if from_tf and os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, TF_WEIGHTS_NAME + ".index")
                ):
3364
                    # Load from a TF 1.0 checkpoint in priority if from_tf
3365
3366
3367
3368
                    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)
                ):
3369
                    # Load from a TF 2.0 checkpoint in priority if from_tf
3370
3371
3372
3373
                    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)
                ):
3374
                    # Load from a Flax checkpoint in priority if from_flax
3375
                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)
3376
                elif use_safetensors is not False and os.path.isfile(
3377
                    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant))
3378
3379
                ):
                    # Load from a safetensors checkpoint
3380
3381
3382
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant)
                    )
3383
                elif use_safetensors is not False and os.path.isfile(
3384
3385
3386
                    os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
                    )
3387
3388
                ):
                    # Load from a sharded safetensors checkpoint
3389
3390
3391
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
                    )
3392
                    is_sharded = True
3393
3394
3395
                elif os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_NAME, variant))
                ):
thomwolf's avatar
thomwolf committed
3396
                    # Load from a PyTorch checkpoint
3397
3398
3399
3400
3401
3402
                    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
3403
                    # Load from a sharded PyTorch checkpoint
3404
3405
3406
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_INDEX_NAME, variant)
                    )
Sylvain Gugger's avatar
Sylvain Gugger committed
3407
                    is_sharded = True
3408
3409
                # At this stage we don't have a weight file so we will raise an error.
                elif os.path.isfile(
3410
3411
                    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)):
3412
                    raise EnvironmentError(
3413
3414
3415
                        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."
3416
                    )
3417
                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)):
3418
                    raise EnvironmentError(
3419
3420
3421
                        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."
3422
                    )
3423
3424
3425
3426
3427
                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
3428
                else:
3429
                    raise EnvironmentError(
3430
3431
                        f"Error no file named {_add_variant(WEIGHTS_NAME, variant)}, {_add_variant(SAFE_WEIGHTS_NAME, variant)},"
                        f" {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME + '.index'} or {FLAX_WEIGHTS_NAME} found in directory"
3432
                        f" {pretrained_model_name_or_path}."
3433
                    )
3434
            elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path)):
3435
                archive_file = pretrained_model_name_or_path
3436
                is_local = True
3437
            elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path + ".index")):
3438
3439
3440
3441
3442
                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."
                    )
3443
                archive_file = os.path.join(subfolder, pretrained_model_name_or_path + ".index")
3444
                is_local = True
3445
            elif is_remote_url(pretrained_model_name_or_path):
3446
                filename = pretrained_model_name_or_path
3447
                resolved_archive_file = download_url(pretrained_model_name_or_path)
3448
            else:
3449
3450
3451
3452
3453
                # set correct filename
                if from_tf:
                    filename = TF2_WEIGHTS_NAME
                elif from_flax:
                    filename = FLAX_WEIGHTS_NAME
3454
                elif use_safetensors is not False:
3455
                    filename = _add_variant(SAFE_WEIGHTS_NAME, variant)
3456
                else:
3457
                    filename = _add_variant(WEIGHTS_NAME, variant)
3458

3459
3460
                try:
                    # Load from URL or cache if already cached
3461
3462
3463
3464
3465
3466
                    cached_file_kwargs = {
                        "cache_dir": cache_dir,
                        "force_download": force_download,
                        "proxies": proxies,
                        "resume_download": resume_download,
                        "local_files_only": local_files_only,
3467
                        "token": token,
3468
3469
3470
                        "user_agent": user_agent,
                        "revision": revision,
                        "subfolder": subfolder,
3471
                        "_raise_exceptions_for_gated_repo": False,
3472
3473
3474
                        "_raise_exceptions_for_missing_entries": False,
                        "_commit_hash": commit_hash,
                    }
3475
                    resolved_archive_file = cached_file(pretrained_model_name_or_path, filename, **cached_file_kwargs)
3476

3477
                    # Since we set _raise_exceptions_for_missing_entries=False, we don't get an exception but a None
3478
                    # result when internet is up, the repo and revision exist, but the file does not.
3479
                    if resolved_archive_file is None and filename == _add_variant(SAFE_WEIGHTS_NAME, variant):
3480
3481
                        # Maybe the checkpoint is sharded, we try to grab the index name in this case.
                        resolved_archive_file = cached_file(
3482
3483
3484
                            pretrained_model_name_or_path,
                            _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant),
                            **cached_file_kwargs,
3485
3486
3487
                        )
                        if resolved_archive_file is not None:
                            is_sharded = True
3488
                        elif use_safetensors:
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
                            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`."
                                )
3501
3502
                        else:
                            # This repo has no safetensors file of any kind, we switch to PyTorch.
3503
                            filename = _add_variant(WEIGHTS_NAME, variant)
3504
                            resolved_archive_file = cached_file(
3505
                                pretrained_model_name_or_path, filename, **cached_file_kwargs
3506
                            )
3507
                    if resolved_archive_file is None and filename == _add_variant(WEIGHTS_NAME, variant):
Sylvain Gugger's avatar
Sylvain Gugger committed
3508
                        # Maybe the checkpoint is sharded, we try to grab the index name in this case.
3509
                        resolved_archive_file = cached_file(
3510
3511
3512
                            pretrained_model_name_or_path,
                            _add_variant(WEIGHTS_INDEX_NAME, variant),
                            **cached_file_kwargs,
3513
                        )
3514
3515
                        if resolved_archive_file is not None:
                            is_sharded = True
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
                    if not local_files_only and not is_offline_mode():
                        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,
3526
3527
                                    "cache_dir": cache_dir,
                                    "local_files_only": local_files_only,
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
                                }
                                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.
3551
3552
3553
3554
                            has_file_kwargs = {
                                "revision": revision,
                                "proxies": proxies,
                                "token": token,
3555
3556
                                "cache_dir": cache_dir,
                                "local_files_only": local_files_only,
3557
                            }
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
                            if has_file(pretrained_model_name_or_path, TF2_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 for TensorFlow weights."
                                    " Use `from_tf=True` to load this model from those weights."
                                )
                            elif has_file(pretrained_model_name_or_path, FLAX_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 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."
                                )
                            else:
                                raise EnvironmentError(
                                    f"{pretrained_model_name_or_path} does not appear to have a file named"
                                    f" {_add_variant(WEIGHTS_NAME, variant)}, {_add_variant(SAFE_WEIGHTS_NAME, variant)},"
                                    f" {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME} or {FLAX_WEIGHTS_NAME}."
                                )
3584

3585
3586
3587
3588
                except EnvironmentError:
                    # Raise any environment error raise by `cached_file`. It will have a helpful error message adapted
                    # to the original exception.
                    raise
3589
                except Exception as e:
3590
                    # For any other exception, we throw a generic error.
3591
                    raise EnvironmentError(
3592
3593
3594
                        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"
3595
3596
                        f" directory containing a file named {_add_variant(WEIGHTS_NAME, variant)},"
                        f" {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME} or {FLAX_WEIGHTS_NAME}."
3597
                    ) from e
3598

3599
            if is_local:
3600
                logger.info(f"loading weights file {archive_file}")
3601
                resolved_archive_file = archive_file
3602
            else:
3603
                logger.info(f"loading weights file {filename} from cache at {resolved_archive_file}")
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
        elif gguf_file:
            from .modeling_gguf_pytorch_utils import load_gguf_checkpoint

            # Case 1: the GGUF file is present locally
            if os.path.isfile(gguf_file):
                gguf_path = gguf_file
            # Case 2: The GGUF path is a location on the Hub
            # Load from URL or cache if already cached
            else:
                cached_file_kwargs = {
                    "cache_dir": cache_dir,
                    "force_download": force_download,
                    "proxies": proxies,
                    "resume_download": resume_download,
                    "local_files_only": local_files_only,
                    "token": token,
                    "user_agent": user_agent,
                    "revision": revision,
                    "subfolder": subfolder,
                    "_raise_exceptions_for_gated_repo": False,
                    "_raise_exceptions_for_missing_entries": False,
                    "_commit_hash": commit_hash,
                }

                gguf_path = cached_file(pretrained_model_name_or_path, gguf_file, **cached_file_kwargs)

            state_dict = load_gguf_checkpoint(gguf_path, return_tensors=True)["tensors"]

            resolved_archive_file = None
            is_sharded = False
3634
        else:
thomwolf's avatar
thomwolf committed
3635
            resolved_archive_file = None
3636

Sylvain Gugger's avatar
Sylvain Gugger committed
3637
3638
        # We'll need to download and cache each checkpoint shard if the checkpoint is sharded.
        if is_sharded:
3639
            # resolved_archive_file becomes a list of files that point to the different checkpoint shards in this case.
Sylvain Gugger's avatar
Sylvain Gugger committed
3640
3641
3642
3643
3644
3645
3646
3647
            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,
3648
                token=token,
Sylvain Gugger's avatar
Sylvain Gugger committed
3649
3650
                user_agent=user_agent,
                revision=revision,
3651
                subfolder=subfolder,
3652
                _commit_hash=commit_hash,
Sylvain Gugger's avatar
Sylvain Gugger committed
3653
3654
            )

3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
        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.")
3671
3672
3673
            elif metadata.get("format") == "mlx":
                # This is a mlx file, we assume weights are compatible with pt
                pass
3674
3675
            else:
                raise ValueError(
3676
                    f"Incompatible safetensors file. File metadata is not ['pt', 'tf', 'flax', 'mlx'] but {metadata.get('format')}"
3677
3678
3679
3680
                )

        from_pt = not (from_tf | from_flax)

3681
3682
        # load pt weights early so that we know which dtype to init the model under
        if from_pt:
3683
            if not is_sharded and state_dict is None:
Sylvain Gugger's avatar
Sylvain Gugger committed
3684
3685
                # Time to load the checkpoint
                state_dict = load_state_dict(resolved_archive_file)
3686

3687
3688
3689
            # 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
3690
            #    weights entry that is of a floating type - we assume all floating dtype weights are of the same dtype
3691
3692
            # we also may have config.torch_dtype available, but we won't rely on it till v5
            dtype_orig = None
3693

3694
3695
3696
            if torch_dtype is not None:
                if isinstance(torch_dtype, str):
                    if torch_dtype == "auto":
3697
3698
3699
                        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
3700
                        else:
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
                            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"
                            )
3713
3714
                    elif hasattr(torch, torch_dtype):
                        torch_dtype = getattr(torch, torch_dtype)
3715
3716
                    else:
                        raise ValueError(
3717
                            f'`torch_dtype` can be one of: `torch.dtype`, `"auto"` or a string of a valid `torch.dtype`, but received {torch_dtype}'
3718
3719
3720
                        )
                dtype_orig = cls._set_default_torch_dtype(torch_dtype)

3721
            # Check if `_keep_in_fp32_modules` is not None
3722
            use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and (
3723
                (torch_dtype == torch.float16) or hasattr(hf_quantizer, "use_keep_in_fp32_modules")
3724
3725
            )

3726
3727
3728
            if is_sharded:
                loaded_state_dict_keys = sharded_metadata["all_checkpoint_keys"]
            else:
3729
                loaded_state_dict_keys = list(state_dict.keys())
3730
3731

            if gguf_path is None and (low_cpu_mem_usage or (use_keep_in_fp32_modules and is_accelerate_available())):
3732
3733
3734
                # 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.
3735
                state_dict = None
3736

3737
3738
        config.name_or_path = pretrained_model_name_or_path

3739
        # Instantiate model.
3740
3741
        init_contexts = [no_init_weights(_enable=_fast_init)]

3742
        if is_deepspeed_zero3_enabled() and not is_quantized:
3743
3744
3745
            import deepspeed

            logger.info("Detected DeepSpeed ZeRO-3: activating zero.init() for this model")
3746
            init_contexts = [deepspeed.zero.Init(config_dict_or_path=deepspeed_config())] + init_contexts
3747
        elif low_cpu_mem_usage:
3748
3749
            init_contexts.append(init_empty_weights())

3750
3751
3752
3753
        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
        )
3754

3755
        with ContextManagers(init_contexts):
3756
            # Let's make sure we don't run the init function of buffer modules
3757
3758
            model = cls(config, *model_args, **model_kwargs)

3759
3760
3761
        # make sure we use the model's config since the __init__ call might have copied it
        config = model.config

3762
3763
        # Check first if we are `from_pt`
        if use_keep_in_fp32_modules:
3764
            if is_accelerate_available() and not is_deepspeed_zero3_enabled():
3765
                low_cpu_mem_usage = True
3766
3767
3768
3769
            keep_in_fp32_modules = model._keep_in_fp32_modules
        else:
            keep_in_fp32_modules = []

3770
3771
3772
        if hf_quantizer is not None:
            hf_quantizer.preprocess_model(
                model=model, device_map=device_map, keep_in_fp32_modules=keep_in_fp32_modules
3773
            )
3774

3775
3776
3777
3778
3779
3780
            # 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

3781
        if isinstance(device_map, str):
3782
            special_dtypes = {}
3783
3784
3785

            if hf_quantizer is not None:
                special_dtypes.update(hf_quantizer.get_special_dtypes_update(model, torch_dtype))
3786
3787
3788
3789
3790
3791
3792
3793
3794

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

3795
3796
            target_dtype = torch_dtype

3797
3798
            if hf_quantizer is not None:
                target_dtype = hf_quantizer.adjust_target_dtype(target_dtype)
3799

Marc Sun's avatar
Marc Sun committed
3800
            no_split_modules = model._get_no_split_modules(device_map)
3801
3802
3803
3804
3805
            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'."
                )
3806

3807
            device_map_kwargs = {"no_split_module_classes": no_split_modules}
3808
            if "special_dtypes" in inspect.signature(infer_auto_device_map).parameters:
3809
                device_map_kwargs["special_dtypes"] = special_dtypes
3810
            elif len(special_dtypes) > 0:
3811
                logger.warning(
3812
3813
3814
                    "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`)."
                )
3815
            if device_map != "sequential":
3816
3817
                max_memory = get_balanced_memory(
                    model,
3818
                    dtype=target_dtype,
3819
                    low_zero=(device_map == "balanced_low_0"),
3820
                    max_memory=max_memory,
3821
                    **device_map_kwargs,
3822
                )
Marc Sun's avatar
Marc Sun committed
3823
3824
            else:
                max_memory = get_max_memory(max_memory)
3825
3826
            if hf_quantizer is not None:
                max_memory = hf_quantizer.adjust_max_memory(max_memory)
3827
            device_map_kwargs["max_memory"] = max_memory
Marc Sun's avatar
Marc Sun committed
3828

3829
3830
            # Make sure tied weights are tied before creating the device map.
            model.tie_weights()
3831
            device_map = infer_auto_device_map(model, dtype=target_dtype, **device_map_kwargs)
3832

3833
3834
            if hf_quantizer is not None:
                hf_quantizer.validate_environment(device_map=device_map)
3835

3836
3837
3838
3839
        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
3840
            check_tied_parameters_on_same_device(tied_params, device_map)
3841

3842
        if from_tf:
3843
            if resolved_archive_file.endswith(".index"):
3844
3845
3846
3847
3848
                # 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:
3849
                    from .modeling_tf_pytorch_utils import load_tf2_checkpoint_in_pytorch_model
3850

Yih-Dar's avatar
Yih-Dar committed
3851
3852
3853
                    model, loading_info = load_tf2_checkpoint_in_pytorch_model(
                        model, resolved_archive_file, allow_missing_keys=True, output_loading_info=True
                    )
3854
                except ImportError:
3855
                    logger.error(
Sylvain Gugger's avatar
Sylvain Gugger committed
3856
3857
3858
                        "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."
3859
                    )
3860
                    raise
3861
3862
3863
3864
3865
3866
3867
        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
3868
3869
3870
                    "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."
3871
3872
                )
                raise
3873
        elif from_pt:
3874
3875
3876
            # restore default dtype
            if dtype_orig is not None:
                torch.set_default_dtype(dtype_orig)
3877

Sylvain Gugger's avatar
Sylvain Gugger committed
3878
3879
3880
3881
3882
3883
3884
3885
            (
                model,
                missing_keys,
                unexpected_keys,
                mismatched_keys,
                offload_index,
                error_msgs,
            ) = cls._load_pretrained_model(
3886
3887
3888
3889
3890
3891
3892
3893
3894
                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,
3895
3896
3897
3898
                device_map=device_map,
                offload_folder=offload_folder,
                offload_state_dict=offload_state_dict,
                dtype=torch_dtype,
3899
                hf_quantizer=hf_quantizer,
3900
                keep_in_fp32_modules=keep_in_fp32_modules,
3901
                gguf_path=gguf_path,
3902
            )
3903

3904
3905
        # make sure token embedding weights are still tied if needed
        model.tie_weights()
3906

3907
        # Set model in evaluation mode to deactivate DropOut modules by default
3908
3909
        model.eval()

3910
        # If it is a model with generation capabilities, attempt to load the generation config
3911
        if model.can_generate() and pretrained_model_name_or_path is not None:
3912
3913
3914
3915
3916
3917
3918
3919
            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,
3920
                    token=token,
3921
3922
3923
3924
3925
3926
                    revision=revision,
                    subfolder=subfolder,
                    _from_auto=from_auto_class,
                    _from_pipeline=from_pipeline,
                    **kwargs,
                )
3927
            except OSError:
3928
3929
3930
3931
3932
                logger.info(
                    "Generation config file not found, using a generation config created from the model config."
                )
                pass

3933
3934
        # Dispatch model with hooks on all devices if necessary
        if device_map is not None:
3935
3936
3937
3938
            device_map_kwargs = {
                "device_map": device_map,
                "offload_dir": offload_folder,
                "offload_index": offload_index,
3939
                "offload_buffers": offload_buffers,
3940
            }
3941
            if "skip_keys" in inspect.signature(dispatch_model).parameters:
3942
                device_map_kwargs["skip_keys"] = model._skip_keys_device_placement
3943
3944
3945
3946
3947
3948
3949
            # For HQQ method we force-set the hooks for single GPU envs
            if (
                "force_hooks" in inspect.signature(dispatch_model).parameters
                and hf_quantizer is not None
                and hf_quantizer.quantization_config.quant_method == QuantizationMethod.HQQ
            ):
                device_map_kwargs["force_hooks"] = True
3950
3951
            if not is_fsdp_enabled() and not is_deepspeed_zero3_enabled():
                dispatch_model(model, **device_map_kwargs)
3952

3953
3954
3955
        if hf_quantizer is not None:
            hf_quantizer.postprocess_model(model)
            model.hf_quantizer = hf_quantizer
Marc Sun's avatar
Marc Sun committed
3956

3957
        if _adapter_model_path is not None:
3958
            model.load_adapter(
3959
                _adapter_model_path,
3960
3961
                adapter_name=adapter_name,
                token=token,
3962
                adapter_kwargs=adapter_kwargs,
3963
3964
            )

thomwolf's avatar
thomwolf committed
3965
        if output_loading_info:
Yih-Dar's avatar
Yih-Dar committed
3966
3967
3968
3969
3970
3971
3972
            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
3973
3974
            return model, loading_info

3975
3976
        return model

3977
    @classmethod
Sylvain Gugger's avatar
Sylvain Gugger committed
3978
3979
3980
3981
    def _load_pretrained_model(
        cls,
        model,
        state_dict,
3982
        loaded_keys,
Sylvain Gugger's avatar
Sylvain Gugger committed
3983
3984
3985
3986
3987
        resolved_archive_file,
        pretrained_model_name_or_path,
        ignore_mismatched_sizes=False,
        sharded_metadata=None,
        _fast_init=True,
3988
        low_cpu_mem_usage=False,
3989
3990
        device_map=None,
        offload_folder=None,
3991
        offload_state_dict=None,
3992
        dtype=None,
3993
        hf_quantizer=None,
3994
        keep_in_fp32_modules=None,
3995
        gguf_path=None,
3996
    ):
Sylvain Gugger's avatar
Sylvain Gugger committed
3997
        is_safetensors = False
3998
        is_quantized = hf_quantizer is not None
3999
4000
        state_dict_folder = None
        state_dict_index = None
4001

Sylvain Gugger's avatar
Sylvain Gugger committed
4002
        if device_map is not None and "disk" in device_map.values():
Sylvain Gugger's avatar
Sylvain Gugger committed
4003
4004
4005
4006
4007
            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
4008
4009
                raise ValueError(
                    "The current `device_map` had weights offloaded to the disk. Please provide an `offload_folder`"
Sylvain Gugger's avatar
Sylvain Gugger committed
4010
4011
                    " 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
4012
                )
Sylvain Gugger's avatar
Sylvain Gugger committed
4013
4014
            if offload_folder is not None:
                os.makedirs(offload_folder, exist_ok=True)
4015
4016
4017
            if offload_state_dict is None:
                offload_state_dict = True

4018
        is_sharded_safetensors = is_safetensors and sharded_metadata is not None
Patrick von Platen's avatar
Patrick von Platen committed
4019
4020
4021
4022

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

4023
        # Retrieve missing & unexpected_keys
4024
4025
        model_state_dict = model.state_dict()
        expected_keys = list(model_state_dict.keys())
4026
4027
        prefix = model.base_model_prefix

Sylvain Gugger's avatar
Sylvain Gugger committed
4028
4029
4030
4031
4032
4033
4034
        def _fix_key(key):
            if "beta" in key:
                return key.replace("beta", "bias")
            if "gamma" in key:
                return key.replace("gamma", "weight")
            return key

4035
        original_loaded_keys = loaded_keys
Sylvain Gugger's avatar
Sylvain Gugger committed
4036
4037
        loaded_keys = [_fix_key(key) for key in loaded_keys]

4038
4039
4040
4041
4042
4043
        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
4044
4045
4046

        # key re-naming operations are never done on the keys
        # that are loaded, but always on the keys of the newly initialized model
4047
4048
        remove_prefix_from_model = not has_prefix_module and expects_prefix_module
        add_prefix_to_model = has_prefix_module and not expects_prefix_module
4049

4050
        if remove_prefix_from_model:
4051
4052
4053
            _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]
4054
        elif add_prefix_to_model:
4055
4056
            expected_keys = [".".join([prefix, s]) for s in expected_keys]

4057
        missing_keys = sorted(set(expected_keys) - set(loaded_keys))
Sylvain Gugger's avatar
Sylvain Gugger committed
4058
        unexpected_keys = set(loaded_keys) - set(expected_keys)
4059

Sylvain Gugger's avatar
Sylvain Gugger committed
4060
4061
4062
4063
4064
4065
4066
        # 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}
4067
        unexpected_keys = sorted(unexpected_keys - model_buffers)
4068

4069
        model.tie_weights()
4070
        if device_map is None and not is_fsdp_enabled() and not is_deepspeed_zero3_enabled():
4071
4072
4073
4074
            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
4075

4076
4077
4078
4079
4080
            # 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
4081
4082

        for group in tied_params:
Sylvain Gugger's avatar
Sylvain Gugger committed
4083
4084
4085
4086
            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
4087
4088
4089
            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]
4090

4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
        # 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]

4101
4102
4103
        if hf_quantizer is not None:
            missing_keys = hf_quantizer.update_missing_keys(model, missing_keys, prefix)

4104
4105
4106
4107
        # 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
4108
4109
                if key in list(model_state_dict.keys()):
                    key = key
4110
4111
                elif f"{prefix}.{key}" in list(model_state_dict.keys()):
                    key = f"{prefix}.{key}"
Susnato Dhar's avatar
Susnato Dhar committed
4112
                elif key.startswith(prefix) and ".".join(key.split(".")[1:]) in list(model_state_dict.keys()):
4113
4114
                    key = ".".join(key.split(".")[1:])
                param = model_state_dict[key]
4115
4116
4117
4118
4119
4120

                # upcast in fp32 if any
                target_dtype = dtype
                if (
                    keep_in_fp32_modules is not None
                    and dtype == torch.float16
4121
4122
4123
                    and any(
                        module_to_keep_in_fp32 in key.split(".") for module_to_keep_in_fp32 in keep_in_fp32_modules
                    )
4124
4125
4126
                ):
                    target_dtype = torch.float32

4127
                if param.device == torch.device("meta"):
4128
                    value = torch.empty(*param.size(), dtype=target_dtype)
4129
                    if (
4130
                        not is_quantized
4131
4132
4133
4134
                        or getattr(hf_quantizer, "requires_parameters_quantization", False)
                        or not hf_quantizer.check_quantized_param(
                            model, param_value=value, param_name=key, state_dict={}
                        )
4135
4136
                    ):
                        set_module_tensor_to_device(model, key, "cpu", value)
4137
                    else:
4138
                        hf_quantizer.create_quantized_param(model, value, key, "cpu", state_dict, unexpected_keys)
4139

4140
        # retrieve uninitialized modules and initialize before maybe overriding that with the pretrained weights.
4141
        if _fast_init:
4142
4143
4144
4145
4146
4147
4148
            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
4149
                not_initialized_submodules = set_initialized_submodules(model, _loaded_keys)
4150
                # If we're about to tie the output embeds to the input embeds we don't need to init them
4151
4152
4153
                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:
4154
4155
4156
                        # 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
4157
4158
            else:
                not_initialized_submodules = dict(model.named_modules())
4159
            # This will only initialize submodules that are not marked as initialized by the line above.
4160
            if is_deepspeed_zero3_enabled() and not is_quantized:
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
                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)
4174

4175
4176
4177
        # Set some modules to fp32 if any
        if keep_in_fp32_modules is not None:
            for name, param in model.named_parameters():
4178
                if any(module_to_keep_in_fp32 in name.split(".") for module_to_keep_in_fp32 in keep_in_fp32_modules):
4179
4180
                    # param = param.to(torch.float32) does not work here as only in the local scope.
                    param.data = param.data.to(torch.float32)
4181

4182
4183
4184
        # Make sure we are able to load base models as well as derived models (with heads)
        start_prefix = ""
        model_to_load = model
4185
        if len(cls.base_model_prefix) > 0 and not hasattr(model, cls.base_model_prefix) and has_prefix_module:
4186
            start_prefix = cls.base_model_prefix + "."
4187
        if len(cls.base_model_prefix) > 0 and hasattr(model, cls.base_model_prefix) and not has_prefix_module:
4188
            model_to_load = getattr(model, cls.base_model_prefix)
Sylvain Gugger's avatar
Sylvain Gugger committed
4189
4190
            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):
4191
                raise ValueError(
4192
                    "The state dictionary of the model you are trying to load is corrupted. Are you sure it was "
4193
4194
                    "properly saved?"
                )
4195
4196
            if device_map is not None:
                device_map = {k.replace(f"{cls.base_model_prefix}.", ""): v for k, v in device_map.items()}
4197

4198
4199
4200
4201
4202
4203
4204
4205
        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
4206
4207
4208
            mismatched_keys = []
            if ignore_mismatched_sizes:
                for checkpoint_key in loaded_keys:
4209
4210
4211
                    # 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
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
                    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
                    ):
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
                        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]
4236
4237
            return mismatched_keys

4238
4239
4240
4241
        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
4242
        if device_map is not None and is_safetensors:
4243
            param_device_map = expand_device_map(device_map, original_loaded_keys, start_prefix)
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
            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
4254
            offload_index = {
4255
                p[len(start_prefix) :]: {"safetensors_file": f, "weight_name": p, "dtype": str_dtype}
4256
                for p, f in weight_map.items()
4257
                if p.startswith(start_prefix) and param_device_map[p[len(start_prefix) :]] == "disk"
Sylvain Gugger's avatar
Sylvain Gugger committed
4258
            }
4259
4260
        else:
            offload_index = None
Sylvain Gugger's avatar
Sylvain Gugger committed
4261

4262
4263
4264
4265
4266
        if state_dict is not None:
            # Whole checkpoint
            mismatched_keys = _find_mismatched_keys(
                state_dict,
                model_state_dict,
4267
                original_loaded_keys,
4268
4269
4270
4271
                add_prefix_to_model,
                remove_prefix_from_model,
                ignore_mismatched_sizes,
            )
4272

4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
            # For GGUF models `state_dict` is never set to None as the state dict is always small
            if gguf_path:
                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,
                    hf_quantizer=hf_quantizer,
                    is_safetensors=is_safetensors,
                    keep_in_fp32_modules=keep_in_fp32_modules,
                    unexpected_keys=unexpected_keys,
                )
            else:
                # Sharded checkpoint or whole but low_cpu_mem_usage==True
4294
4295
4296
4297
4298
4299
                assign_to_params_buffers = check_support_param_buffer_assignment(
                    model_to_load, state_dict, start_prefix
                )
                error_msgs = _load_state_dict_into_model(
                    model_to_load, state_dict, start_prefix, assign_to_params_buffers
                )
4300
4301

        else:
Sylvain Gugger's avatar
Sylvain Gugger committed
4302
4303
4304
4305
4306
            # 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 = []
4307
            mismatched_keys = []
Sylvain Gugger's avatar
Sylvain Gugger committed
4308
4309
            if not is_safetensors:
                offload_index = {} if device_map is not None and "disk" in device_map.values() else None
4310
4311
4312
4313
4314
4315
4316
            if offload_state_dict:
                state_dict_folder = tempfile.mkdtemp()
                state_dict_index = {}
            else:
                state_dict_folder = None
                state_dict_index = None

4317
            if is_sharded_safetensors:
4318
4319
4320
                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
4321
4322
4323
4324
                disk_only_shard_files = [os.path.join(folder, f) for f in disk_only_shard_files]
            else:
                disk_only_shard_files = []

4325
4326
            if len(resolved_archive_file) > 1:
                resolved_archive_file = logging.tqdm(resolved_archive_file, desc="Loading checkpoint shards")
4327
            assign_to_params_buffers = None
Sylvain Gugger's avatar
Sylvain Gugger committed
4328
            for shard_file in resolved_archive_file:
Sylvain Gugger's avatar
Sylvain Gugger committed
4329
4330
4331
                # 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
4332
                state_dict = load_state_dict(shard_file, is_quantized=is_quantized)
4333

Sylvain Gugger's avatar
Sylvain Gugger committed
4334
4335
                # Mistmatched keys contains tuples key/shape1/shape2 of weights in the checkpoint that have a shape not
                # matching the weights in the model.
4336
4337
4338
                mismatched_keys += _find_mismatched_keys(
                    state_dict,
                    model_state_dict,
4339
                    original_loaded_keys,
4340
4341
4342
4343
                    add_prefix_to_model,
                    remove_prefix_from_model,
                    ignore_mismatched_sizes,
                )
4344
                if low_cpu_mem_usage:
4345
                    if is_fsdp_enabled() and not is_local_dist_rank_0() and not is_quantized:
4346
4347
                        for key, param in model_to_load.state_dict().items():
                            if param.device == torch.device("meta"):
4348
4349
4350
                                set_module_tensor_to_device(
                                    model_to_load, key, "cpu", torch.empty(*param.size(), dtype=dtype)
                                )
4351
                    else:
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
                        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,
4364
                            hf_quantizer=hf_quantizer,
4365
4366
                            is_safetensors=is_safetensors,
                            keep_in_fp32_modules=keep_in_fp32_modules,
4367
                            unexpected_keys=unexpected_keys,
4368
4369
                        )
                        error_msgs += new_error_msgs
4370
                else:
4371
4372
4373
4374
4375
4376
4377
4378
                    # Sharded checkpoint or whole but low_cpu_mem_usage==True
                    if assign_to_params_buffers is None:
                        assign_to_params_buffers = check_support_param_buffer_assignment(
                            model_to_load, state_dict, start_prefix
                        )
                    error_msgs += _load_state_dict_into_model(
                        model_to_load, state_dict, start_prefix, assign_to_params_buffers
                    )
4379

4380
4381
4382
4383
                # force memory release
                del state_dict
                gc.collect()

4384
            if offload_index is not None and len(offload_index) > 0:
Sylvain Gugger's avatar
Sylvain Gugger committed
4385
4386
4387
                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
4388
4389
4390
4391
4392
4393
                    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
4394
                    offload_index = {f"{prefix}.{key}": value for key, value in offload_index.items()}
Sylvain Gugger's avatar
Sylvain Gugger committed
4395
4396
4397
                if not is_safetensors:
                    save_offload_index(offload_index, offload_folder)
                    offload_index = None
4398
4399
4400

            if offload_state_dict:
                # Load back temporarily offloaded state dict
4401
                load_offloaded_weights(model_to_load, state_dict_index, state_dict_folder)
4402
4403
                shutil.rmtree(state_dict_folder)

4404
4405
        if len(error_msgs) > 0:
            error_msg = "\n\t".join(error_msgs)
4406
4407
4408
4409
            if "size mismatch" in error_msg:
                error_msg += (
                    "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
                )
4410
4411
            raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")

4412
        if len(unexpected_keys) > 0:
Sylvain Gugger's avatar
Sylvain Gugger committed
4413
            archs = [] if model.config.architectures is None else model.config.architectures
4414
            warner = logger.warning if model.__class__.__name__ in archs else logger.info
Sylvain Gugger's avatar
Sylvain Gugger committed
4415
            warner(
Sylvain Gugger's avatar
Sylvain Gugger committed
4416
4417
4418
4419
4420
4421
4422
                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)."
4423
4424
4425
4426
4427
            )
        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
4428
4429
4430
                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."
4431
            )
4432
        elif len(mismatched_keys) == 0:
4433
            logger.info(
Sylvain Gugger's avatar
Sylvain Gugger committed
4434
4435
4436
4437
                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."
4438
            )
4439
4440
4441
4442
4443
4444
4445
4446
        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
4447
4448
4449
4450
                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."
4451
            )
4452

Sylvain Gugger's avatar
Sylvain Gugger committed
4453
        return model, missing_keys, unexpected_keys, mismatched_keys, offload_index, error_msgs
4454
4455

    def retrieve_modules_from_names(self, names, add_prefix=False, remove_prefix=False):
4456
        module_keys = {".".join(key.split(".")[:-1]) for key in names}
4457

Patrick von Platen's avatar
Patrick von Platen committed
4458
4459
        # torch.nn.ParameterList is a special case where two parameter keywords
        # are appended to the module name, *e.g.* bert.special_embeddings.0
4460
        module_keys = module_keys.union(
4461
            {".".join(key.split(".")[:-2]) for key in names if len(key) > 0 and key[-1].isdigit()}
4462
        )
Patrick von Platen's avatar
Patrick von Platen committed
4463

4464
4465
4466
4467
        retrieved_modules = []
        # retrieve all modules that has at least one missing weight name
        for name, module in self.named_modules():
            if remove_prefix:
4468
4469
                _prefix = f"{self.base_model_prefix}."
                name = name[len(_prefix) :] if name.startswith(_prefix) else name
4470
            elif add_prefix:
Patrick von Platen's avatar
Patrick von Platen committed
4471
                name = ".".join([self.base_model_prefix, name]) if len(name) > 0 else self.base_model_prefix
4472
4473
4474
4475
4476
4477

            if name in module_keys:
                retrieved_modules.append(module)

        return retrieved_modules

4478
    @staticmethod
4479
4480
4481
    def _load_pretrained_model_low_mem(
        model, loaded_state_dict_keys, resolved_archive_file, start_prefix="", hf_quantizer=None
    ):
4482
4483
4484
        """
        This is an experimental function that loads the model using ~1.x model size CPU memory

4485
        Before you call it do:
4486

4487
        1. save which state_dict keys are available
4488
4489
4490
4491
4492
4493
4494
4495
        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

4496
4497
        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.
4498
4499
        """

4500
4501
        _move_model_to_meta(model, loaded_state_dict_keys, start_prefix)
        state_dict = load_state_dict(resolved_archive_file)
4502
4503
4504
4505
4506
4507
4508
4509
4510
        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,
        )
4511
        return error_msgs
4512

4513
4514
4515
4516
4517
4518
    @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.

4519
4520
4521
4522
4523
4524
        <Tip warning={true}>

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

        </Tip>

4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
        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

4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
    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)

4589
4590
4591
4592
    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.
        """
4593
4594

        # Skip the check during tracing.
4595
        if is_torch_fx_proxy(input_ids) or torch.jit.is_tracing() or is_torchdynamo_compiling():
4596
4597
            return

4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
        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)

4624
4625
    @property
    def _is_quantized_training_enabled(self):
4626
        warnings.warn(
4627
4628
4629
4630
4631
4632
4633
4634
4635
            "`_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
4636

4637
PreTrainedModel.push_to_hub = copy_func(PreTrainedModel.push_to_hub)
4638
4639
4640
4641
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"
    )
4642
4643


thomwolf's avatar
thomwolf committed
4644
class PoolerStartLogits(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4645
4646
    """
    Compute SQuAD start logits from sequence hidden states.
4647

Sylvain Gugger's avatar
Sylvain Gugger committed
4648
    Args:
4649
4650
        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
4651
4652
4653
    """

    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
4654
        super().__init__()
thomwolf's avatar
thomwolf committed
4655
4656
        self.dense = nn.Linear(config.hidden_size, 1)

Sylvain Gugger's avatar
Sylvain Gugger committed
4657
4658
4659
4660
4661
    def forward(
        self, hidden_states: torch.FloatTensor, p_mask: Optional[torch.FloatTensor] = None
    ) -> torch.FloatTensor:
        """
        Args:
4662
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4663
                The final hidden states of the model.
4664
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4665
4666
                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
4667
4668

        Returns:
4669
            `torch.FloatTensor`: The start logits for SQuAD.
thomwolf's avatar
thomwolf committed
4670
        """
thomwolf's avatar
thomwolf committed
4671
4672
4673
        x = self.dense(hidden_states).squeeze(-1)

        if p_mask is not None:
Lysandre Debut's avatar
Lysandre Debut committed
4674
            if get_parameter_dtype(self) == torch.float16:
4675
4676
4677
                x = x * (1 - p_mask) - 65500 * p_mask
            else:
                x = x * (1 - p_mask) - 1e30 * p_mask
thomwolf's avatar
thomwolf committed
4678
4679
4680
4681
4682
4683

        return x


class PoolerEndLogits(nn.Module):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
4684
    Compute SQuAD end logits from sequence hidden states.
4685

Sylvain Gugger's avatar
Sylvain Gugger committed
4686
    Args:
4687
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
4688
4689
            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
4690
4691
4692
    """

    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
4693
        super().__init__()
thomwolf's avatar
thomwolf committed
4694
4695
4696
4697
4698
        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
4699
4700
4701
4702
4703
4704
4705
4706
4707
    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:
4708
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4709
                The final hidden states of the model.
4710
            start_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4711
                The hidden states of the first tokens for the labeled span.
4712
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4713
                The position of the first token for the labeled span.
4714
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4715
4716
                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
4717

4718
        <Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
4719

Stas Bekman's avatar
Stas Bekman committed
4720
4721
        One of `start_states` or `start_positions` should be not `None`. If both are set, `start_positions` overrides
        `start_states`.
4722
4723

        </Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
4724
4725

        Returns:
4726
            `torch.FloatTensor`: The end logits for SQuAD.
thomwolf's avatar
thomwolf committed
4727
        """
4728
4729
4730
        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
4731
        if start_positions is not None:
4732
            slen, hsz = hidden_states.shape[-2:]
4733
4734
4735
            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
4736
4737
4738
4739
4740
4741
4742

        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
4743
            if get_parameter_dtype(self) == torch.float16:
4744
4745
4746
                x = x * (1 - p_mask) - 65500 * p_mask
            else:
                x = x * (1 - p_mask) - 1e30 * p_mask
thomwolf's avatar
thomwolf committed
4747
4748
4749
4750
4751

        return x


class PoolerAnswerClass(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4752
4753
4754
4755
    """
    Compute SQuAD 2.0 answer class from classification and start tokens hidden states.

    Args:
4756
4757
        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
4758
    """
4759

thomwolf's avatar
thomwolf committed
4760
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
4761
        super().__init__()
thomwolf's avatar
thomwolf committed
4762
4763
4764
4765
        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
4766
4767
4768
4769
4770
4771
4772
    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:
4773
4774
        """
        Args:
4775
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4776
                The final hidden states of the model.
4777
            start_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4778
                The hidden states of the first tokens for the labeled span.
4779
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4780
                The position of the first token for the labeled span.
4781
4782
4783
4784
            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
4785

Stas Bekman's avatar
Stas Bekman committed
4786
4787
        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
4788

4789
        </Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
4790
4791

        Returns:
4792
            `torch.FloatTensor`: The SQuAD 2.0 answer class.
thomwolf's avatar
thomwolf committed
4793
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
4794
        # No dependency on end_feature so that we can obtain one single `cls_logits` for each sample.
4795
        hsz = hidden_states.shape[-1]
4796
4797
4798
        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
4799
        if start_positions is not None:
4800
4801
            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
4802
4803

        if cls_index is not None:
4804
4805
            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
4806
        else:
4807
            cls_token_state = hidden_states[:, -1, :]  # shape (bsz, hsz)
thomwolf's avatar
thomwolf committed
4808
4809
4810
4811
4812
4813
4814
4815

        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


4816
4817
4818
@dataclass
class SquadHeadOutput(ModelOutput):
    """
4819
    Base class for outputs of question answering models using a [`~modeling_utils.SQuADHead`].
4820
4821

    Args:
4822
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned if both `start_positions` and `end_positions` are provided):
Sylvain Gugger's avatar
Sylvain Gugger committed
4823
4824
            Classification loss as the sum of start token, end token (and is_impossible if provided) classification
            losses.
4825
        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):
4826
            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
4827
        start_top_index (`torch.LongTensor` of shape `(batch_size, config.start_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
4828
            Indices for the top config.start_n_top start token possibilities (beam-search).
4829
4830
        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
4831
            (beam-search).
4832
4833
4834
4835
        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.
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846

    """

    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
4847
class SQuADHead(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4848
4849
    r"""
    A SQuAD head inspired by XLNet.
4850

Sylvain Gugger's avatar
Sylvain Gugger committed
4851
    Args:
4852
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
4853
4854
            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
4855
    """
4856

thomwolf's avatar
thomwolf committed
4857
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
4858
        super().__init__()
thomwolf's avatar
thomwolf committed
4859
4860
4861
4862
4863
4864
4865
        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
4866
    @replace_return_docstrings(output_type=SquadHeadOutput, config_class=PretrainedConfig)
4867
    def forward(
4868
        self,
Sylvain Gugger's avatar
Sylvain Gugger committed
4869
4870
4871
4872
4873
4874
        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,
4875
        return_dict: bool = False,
Sylvain Gugger's avatar
Sylvain Gugger committed
4876
4877
    ) -> Union[SquadHeadOutput, Tuple[torch.FloatTensor]]:
        """
Lysandre's avatar
Lysandre committed
4878
        Args:
4879
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Lysandre's avatar
Lysandre committed
4880
                Final hidden states of the model on the sequence tokens.
4881
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Lysandre's avatar
Lysandre committed
4882
                Positions of the first token for the labeled span.
4883
            end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Lysandre's avatar
Lysandre committed
4884
                Positions of the last token for the labeled span.
4885
4886
4887
            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
4888
                Whether the question has a possible answer in the paragraph or not.
4889
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4890
4891
                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
                should be masked.
4892
            return_dict (`bool`, *optional*, defaults to `False`):
4893
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
Sylvain Gugger's avatar
Sylvain Gugger committed
4894

Lysandre's avatar
Lysandre committed
4895
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
4896
        """
thomwolf's avatar
thomwolf committed
4897
        start_logits = self.start_logits(hidden_states, p_mask=p_mask)
thomwolf's avatar
thomwolf committed
4898
4899
4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920

        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
4921

4922
            return SquadHeadOutput(loss=total_loss) if return_dict else (total_loss,)
thomwolf's avatar
thomwolf committed
4923
4924
4925
4926

        else:
            # during inference, compute the end logits based on beam search
            bsz, slen, hsz = hidden_states.size()
4927
            start_log_probs = nn.functional.softmax(start_logits, dim=-1)  # shape (bsz, slen)
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938

            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
4939
4940
            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)
4941
            end_log_probs = nn.functional.softmax(end_logits, dim=1)  # shape (bsz, slen, start_n_top)
thomwolf's avatar
thomwolf committed
4942

4943
4944
4945
            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
4946
4947
4948
4949
4950
4951
            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)

4952
            if not return_dict:
4953
4954
4955
4956
4957
4958
4959
4960
4961
                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
4962
4963
4964


class SequenceSummary(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4965
4966
4967
4968
    r"""
    Compute a single vector summary of a sequence hidden states.

    Args:
4969
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
4970
4971
            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
4972

4973
            - **summary_type** (`str`) -- The method to use to make this summary. Accepted values are:
Sylvain Gugger's avatar
Sylvain Gugger committed
4974

4975
4976
4977
4978
4979
                - `"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
4980

4981
            - **summary_use_proj** (`bool`) -- Add a projection after the vector extraction.
Sylvain Gugger's avatar
Sylvain Gugger committed
4982
4983
4984
4985
4986
4987
            - **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
4988
    """
4989

4990
    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
4991
        super().__init__()
thomwolf's avatar
thomwolf committed
4992

4993
        self.summary_type = getattr(config, "summary_type", "last")
4994
        if self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
4995
4996
4997
4998
4999
            # 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
5000
        self.summary = Identity()
5001
5002
        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:
5003
                num_classes = config.num_labels
thomwolf's avatar
thomwolf committed
5004
5005
5006
5007
            else:
                num_classes = config.hidden_size
            self.summary = nn.Linear(config.hidden_size, num_classes)

5008
        activation_string = getattr(config, "summary_activation", None)
Lysandre's avatar
Lysandre committed
5009
        self.activation: Callable = get_activation(activation_string) if activation_string else Identity()
thomwolf's avatar
thomwolf committed
5010

thomwolf's avatar
thomwolf committed
5011
        self.first_dropout = Identity()
5012
        if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
5013
5014
            self.first_dropout = nn.Dropout(config.summary_first_dropout)

thomwolf's avatar
thomwolf committed
5015
        self.last_dropout = Identity()
5016
        if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
5017
            self.last_dropout = nn.Dropout(config.summary_last_dropout)
thomwolf's avatar
thomwolf committed
5018

Sylvain Gugger's avatar
Sylvain Gugger committed
5019
5020
5021
5022
5023
5024
5025
    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:
5026
            hidden_states (`torch.FloatTensor` of shape `[batch_size, seq_len, hidden_size]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
5027
                The hidden states of the last layer.
5028
            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
5029
                Used if `summary_type == "cls_index"` and takes the last token of the sequence as classification token.
Sylvain Gugger's avatar
Sylvain Gugger committed
5030
5031

        Returns:
5032
            `torch.FloatTensor`: The summary of the sequence hidden states.
thomwolf's avatar
thomwolf committed
5033
        """
5034
        if self.summary_type == "last":
thomwolf's avatar
thomwolf committed
5035
            output = hidden_states[:, -1]
5036
        elif self.summary_type == "first":
thomwolf's avatar
thomwolf committed
5037
            output = hidden_states[:, 0]
5038
        elif self.summary_type == "mean":
thomwolf's avatar
thomwolf committed
5039
            output = hidden_states.mean(dim=1)
5040
        elif self.summary_type == "cls_index":
thomwolf's avatar
thomwolf committed
5041
            if cls_index is None:
Lysandre's avatar
Lysandre committed
5042
5043
5044
5045
5046
                cls_index = torch.full_like(
                    hidden_states[..., :1, :],
                    hidden_states.shape[-2] - 1,
                    dtype=torch.long,
                )
thomwolf's avatar
thomwolf committed
5047
            else:
thomwolf's avatar
thomwolf committed
5048
                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
5049
                cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
thomwolf's avatar
thomwolf committed
5050
            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
5051
5052
            output = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, XX, hidden_size)
        elif self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
5053
5054
            raise NotImplementedError

5055
        output = self.first_dropout(output)
thomwolf's avatar
thomwolf committed
5056
5057
        output = self.summary(output)
        output = self.activation(output)
5058
        output = self.last_dropout(output)
thomwolf's avatar
thomwolf committed
5059
5060
5061
5062

        return output


5063
def unwrap_model(model: nn.Module, recursive: bool = False) -> nn.Module:
5064
5065
5066
5067
    """
    Recursively unwraps a model from potential containers (as used in distributed training).

    Args:
5068
        model (`torch.nn.Module`): The model to unwrap.
5069
5070
5071
        recursive (`bool`, *optional*, defaults to `False`):
            Whether to recursively extract all cases of `module.module` from `model` as well as unwrap child sublayers
            recursively, not just the top-level distributed containers.
5072
    """
5073
5074
5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
    # Use accelerate implementation if available (should always be the case when using torch)
    # This is for pytorch, as we also have to handle things like dynamo
    if is_accelerate_available():
        kwargs = {}
        if recursive:
            if not is_accelerate_available("0.29.0"):
                raise RuntimeError(
                    "Setting `recursive=True` to `unwrap_model` requires `accelerate` v0.29.0. Please upgrade your version of accelerate"
                )
            else:
                kwargs["recursive"] = recursive
        return extract_model_from_parallel(model, **kwargs)
5085
    else:
5086
5087
5088
5089
5090
        # 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
5091
5092


5093
def expand_device_map(device_map, param_names, start_prefix):
Sylvain Gugger's avatar
Sylvain Gugger committed
5094
5095
5096
5097
    """
    Expand a device map to return the correspondance parameter name to device.
    """
    new_device_map = {}
5098
    param_names = [p[len(start_prefix) :] for p in param_names if p.startswith(start_prefix)]
Sylvain Gugger's avatar
Sylvain Gugger committed
5099
    for module, device in device_map.items():
5100
5101
5102
        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
5103
5104
5105
    return new_device_map


5106
def get_disk_only_shard_files(device_map, sharded_metadata, start_prefix):
Sylvain Gugger's avatar
Sylvain Gugger committed
5107
5108
5109
    """
    Returns the list of shard files containing only weights offloaded to disk.
    """
5110
5111
5112
5113

    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
5114
    files_content = collections.defaultdict(list)
5115
    for weight_name, filename in weight_map.items():
Sylvain Gugger's avatar
Sylvain Gugger committed
5116
5117
5118
5119
5120
        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"}]