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

import torch
34
from packaging import version
Sylvain Gugger's avatar
Sylvain Gugger committed
35
from torch import Tensor, nn
36
from torch.nn import CrossEntropyLoss, Identity
37
from torch.utils.checkpoint import checkpoint
38

39
from .activations import get_activation
40
from .configuration_utils import PretrainedConfig
41
from .dynamic_module_utils import custom_object_save
42
from .generation import GenerationConfig, GenerationMixin
43
from .integrations import PeftAdapterMixin, deepspeed_config, is_deepspeed_zero3_enabled
44
45
46
47
from .pytorch_utils import (  # noqa: F401
    Conv1D,
    apply_chunking_to_forward,
    find_pruneable_heads_and_indices,
Thomas Wang's avatar
Thomas Wang committed
48
    id_tensor_storage,
49
50
51
52
    prune_conv1d_layer,
    prune_layer,
    prune_linear_layer,
)
53
from .safetensors_conversion import auto_conversion
54
from .utils import (
55
56
    ADAPTER_SAFE_WEIGHTS_NAME,
    ADAPTER_WEIGHTS_NAME,
57
    CONFIG_NAME,
Aymeric Augustin's avatar
Aymeric Augustin committed
58
    DUMMY_INPUTS,
59
    FLAX_WEIGHTS_NAME,
60
61
    SAFE_WEIGHTS_INDEX_NAME,
    SAFE_WEIGHTS_NAME,
62
63
    TF2_WEIGHTS_NAME,
    TF_WEIGHTS_NAME,
Sylvain Gugger's avatar
Sylvain Gugger committed
64
    WEIGHTS_INDEX_NAME,
65
    WEIGHTS_NAME,
66
    ContextManagers,
67
    ModelOutput,
Sylvain Gugger's avatar
Sylvain Gugger committed
68
    PushToHubMixin,
69
    cached_file,
70
    copy_func,
71
    download_url,
72
    extract_commit_hash,
73
    has_file,
74
    is_accelerate_available,
75
    is_auto_awq_available,
Marc Sun's avatar
Marc Sun committed
76
    is_auto_gptq_available,
77
    is_bitsandbytes_available,
78
    is_flash_attn_2_available,
79
    is_offline_mode,
80
    is_optimum_available,
81
    is_peft_available,
82
    is_remote_url,
83
    is_safetensors_available,
84
    is_torch_sdpa_available,
85
    is_torch_tpu_available,
86
    logging,
Sylvain Gugger's avatar
Sylvain Gugger committed
87
    replace_return_docstrings,
88
    strtobool,
89
)
90
from .utils.hub import convert_file_size_to_int, get_checkpoint_shard_files
91
92
93
94
95
96
from .utils.import_utils import (
    ENV_VARS_TRUE_VALUES,
    is_sagemaker_mp_enabled,
    is_torch_fx_proxy,
    is_torchdynamo_compiling,
)
97
from .utils.quantization_config import AwqConfig, BitsAndBytesConfig, GPTQConfig, QuantizationMethod
98
from .utils.versions import require_version_core
99

Aymeric Augustin's avatar
Aymeric Augustin committed
100

101
102
103
XLA_USE_BF16 = os.environ.get("XLA_USE_BF16", "0").upper()
XLA_DOWNCAST_BF16 = os.environ.get("XLA_DOWNCAST_BF16", "0").upper()

104
105
if is_accelerate_available():
    from accelerate import dispatch_model, infer_auto_device_map, init_empty_weights
106
    from accelerate.hooks import add_hook_to_module
107
    from accelerate.utils import (
108
        check_tied_parameters_on_same_device,
109
        find_tied_parameters,
110
        get_balanced_memory,
Marc Sun's avatar
Marc Sun committed
111
        get_max_memory,
112
113
114
115
116
117
        load_offloaded_weights,
        offload_weight,
        save_offload_index,
        set_module_tensor_to_device,
    )

118
119
120
121
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
122

Lysandre Debut's avatar
Lysandre Debut committed
123
logger = logging.get_logger(__name__)
124

125
126
127
128

_init_weights = True


129
def is_fsdp_enabled():
130
131
132
133
    return (
        torch.distributed.is_available()
        and torch.distributed.is_initialized()
        and strtobool(os.environ.get("ACCELERATE_USE_FSDP", "False")) == 1
134
        and strtobool(os.environ.get("FSDP_CPU_RAM_EFFICIENT_LOADING", "False")) == 1
135
    )
136
137


138
139
140
141
142
143
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
    )
144
145


146
147
148
149
150
151
152
153
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

154
155
156
if is_peft_available():
    from .utils import find_adapter_config_file

157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
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,
}

174

175
176
177
178
179
180
181
182
@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
183
    old_init_weights = _init_weights
184

185
186
    if _enable:
        _init_weights = False
187
188
189
190
191
192
193

        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)
194
195
196
    try:
        yield
    finally:
197
        _init_weights = old_init_weights
198
199
200
201
        if _enable:
            # # Restore the original initialization functions
            for name, init_func in TORCH_INIT_FUNCTIONS.items():
                setattr(torch.nn.init, name, init_func)
202
203


Lysandre Debut's avatar
Lysandre Debut committed
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
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


219
220
221
222
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
223
224
225
    try:
        return next(parameter.parameters()).dtype
    except StopIteration:
Sylvain Gugger's avatar
Sylvain Gugger committed
226
        # For nn.DataParallel compatibility in PyTorch > 1.5
Lysandre Debut's avatar
Lysandre Debut committed
227
228
229
230
231
232
233
234
235
236

        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


237
238
def get_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
239
    Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.
240
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
241
242
243
244
    last_dtype = None
    for t in parameter.parameters():
        last_dtype = t.dtype
        if t.is_floating_point():
245
246
247
            # 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
248
249
250
251
252
            # NOTE: `is_torch_tpu_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_tpu_available():
                return torch.bfloat16
            if XLA_DOWNCAST_BF16 in ENV_VARS_TRUE_VALUES and is_torch_tpu_available():
                if t.dtype == torch.float:
253
                    return torch.bfloat16
254
255
                if t.dtype == torch.double:
                    return torch.float32
Sylvain Gugger's avatar
Sylvain Gugger committed
256
            return t.dtype
257

Sylvain Gugger's avatar
Sylvain Gugger committed
258
259
260
    if last_dtype is not None:
        # if no floating dtype was found return whatever the first dtype is
        return last_dtype
261

262
263
264
265
266
267
268
269
270
271
272
273
274
    # 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:
275
276
        # fallback to the last dtype
        return last_tuple[1].dtype
277

278
279
280
281
282
283
284
    # fallback to buffer dtype
    for t in parameter.buffers():
        last_dtype = t.dtype
        if t.is_floating_point():
            return t.dtype
    return last_dtype

285
286
287
288
289
290
291
292
293
294
295
296
297
298

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
299
    Returns the first found floating dtype in `state_dict` if there is one, otherwise returns the first dtype.
300
301
302
303
304
305
306
    """
    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
307
        return next(state_dict.values()).dtype
308
309


Sylvain Gugger's avatar
Sylvain Gugger committed
310
311
312
313
314
315
316
317
318
319
320
321
322
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
323
    bit_search = re.search(r"[^\d](\d+)$", str(dtype))
Sylvain Gugger's avatar
Sylvain Gugger committed
324
325
326
327
328
329
    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


330
331
332
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
333
334
335
336
337
338
339
340
341
342
343
    """
    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
344
    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
345
346
347
348
349
350
351
352
353
    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"`).
354
355
        weights_name (`str`, *optional*, defaults to `"pytorch_model.bin"`):
            The name of the model save file.
Sylvain Gugger's avatar
Sylvain Gugger committed
356
357
358
    """
    max_shard_size = convert_file_size_to_int(max_shard_size)

Thomas Wang's avatar
Thomas Wang committed
359
360
    sharded_state_dicts = [{}]
    last_block_size = 0
Sylvain Gugger's avatar
Sylvain Gugger committed
361
    total_size = 0
Thomas Wang's avatar
Thomas Wang committed
362
    storage_id_to_block = {}
Sylvain Gugger's avatar
Sylvain Gugger committed
363
364

    for key, weight in state_dict.items():
365
366
367
368
369
370
        # 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
371
372
373
374
375
376
377

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

Sylvain Gugger's avatar
Sylvain Gugger committed
378
379
        weight_size = weight.numel() * dtype_byte_size(weight.dtype)

Sylvain Gugger's avatar
Sylvain Gugger committed
380
381
382
        # 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
383
384
            sharded_state_dicts.append({})
            last_block_size = 0
Sylvain Gugger's avatar
Sylvain Gugger committed
385

Thomas Wang's avatar
Thomas Wang committed
386
387
        sharded_state_dicts[-1][key] = weight
        last_block_size += weight_size
Sylvain Gugger's avatar
Sylvain Gugger committed
388
        total_size += weight_size
Thomas Wang's avatar
Thomas Wang committed
389
        storage_id_to_block[storage_id] = len(sharded_state_dicts) - 1
Sylvain Gugger's avatar
Sylvain Gugger committed
390
391
392

    # If we only have one shard, we return it
    if len(sharded_state_dicts) == 1:
393
        return {weights_name: sharded_state_dicts[0]}, None
Sylvain Gugger's avatar
Sylvain Gugger committed
394
395
396
397
398

    # Otherwise, let's build the index
    weight_map = {}
    shards = {}
    for idx, shard in enumerate(sharded_state_dicts):
399
400
401
402
        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
403
404
405
406
407
408
409
410
411
412
        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


413
def load_sharded_checkpoint(model, folder, strict=True, prefer_safe=True):
414
415
416
417
418
419
420
421
422
423
424
425
426
    """
    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.
427
428
429
        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.
430
431
432
433
434
435
436
437

    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)
438
    safe_index_file = os.path.join(folder, SAFE_WEIGHTS_INDEX_NAME)
439

440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
    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:
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
        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)

483
    loader = safe_load_file if load_safe else partial(torch.load, map_location="cpu", weights_only=True)
484

485
    for shard_file in shard_files:
486
        state_dict = loader(os.path.join(folder, shard_file))
487
488
        model.load_state_dict(state_dict, strict=False)

489
        # Make sure memory is freed before we load the next state dict.
490
491
492
493
494
495
496
        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)


Sylvain Gugger's avatar
Sylvain Gugger committed
497
498
499
500
def load_state_dict(checkpoint_file: Union[str, os.PathLike]):
    """
    Reads a PyTorch checkpoint file, returning properly formatted errors if they arise.
    """
501
502
503
504
505
506
507
508
509
510
    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()
        if metadata.get("format") not in ["pt", "tf", "flax"]:
            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
511
    try:
512
        if (
513
514
            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()):
515
516
517
            map_location = "meta"
        else:
            map_location = "cpu"
518

519
        return torch.load(checkpoint_file, map_location=map_location, weights_only=True)
Sylvain Gugger's avatar
Sylvain Gugger committed
520
521
522
    except Exception as e:
        try:
            with open(checkpoint_file) as f:
523
                if f.read(7) == "version":
Sylvain Gugger's avatar
Sylvain Gugger committed
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
                    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."
            )


542
543
544
545
546
547
548
549
550
551
552
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.
    """
    for module_name, module in model.named_modules():
        loaded_keys = [k.replace(f"{module_name}.", "") for k in state_dict_keys if k.startswith(f"{module_name}.")]
        if len(set(module.state_dict().keys()) - set(loaded_keys)) == 0:
            module._is_hf_initialized = True


Sylvain Gugger's avatar
Sylvain Gugger committed
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
def _load_state_dict_into_model(model_to_load, state_dict, start_prefix):
    # Convert old format to new format if needed from a PyTorch state_dict
    old_keys = []
    new_keys = []
    for key in state_dict.keys():
        new_key = None
        if "gamma" in key:
            new_key = key.replace("gamma", "weight")
        if "beta" in key:
            new_key = key.replace("beta", "bias")
        if new_key:
            old_keys.append(key)
            new_keys.append(new_key)
    for old_key, new_key in zip(old_keys, new_keys):
        state_dict[new_key] = state_dict.pop(old_key)

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

    error_msgs = []

    # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants
    # so we need to apply the function recursively.
579
    def load(module: nn.Module, state_dict, prefix=""):
Sylvain Gugger's avatar
Sylvain Gugger committed
580
581
        local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
        args = (state_dict, prefix, local_metadata, True, [], [], error_msgs)
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
        # 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
601
602
603

        for name, child in module._modules.items():
            if child is not None:
604
                load(child, state_dict, prefix + name + ".")
Sylvain Gugger's avatar
Sylvain Gugger committed
605

606
607
608
609
    load(model_to_load, state_dict, prefix=start_prefix)
    # Delete `state_dict` so it could be collected by GC earlier. Note that `state_dict` is a copy of the argument, so
    # it's safe to delete it.
    del state_dict
Sylvain Gugger's avatar
Sylvain Gugger committed
610
611
612
613

    return error_msgs


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
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)


663
664
665
666
667
668
669
670
671
672
673
674
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,
675
    is_quantized=False,
Sylvain Gugger's avatar
Sylvain Gugger committed
676
    is_safetensors=False,
677
    keep_in_fp32_modules=None,
678
):
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
    """
    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.

696
    if is_quantized:
697
        from .integrations import set_module_quantized_tensor_to_device
698

699
700
    error_msgs = []

701
702
703
704
705
706
707
708
709
710
711
712
713
    old_keys = []
    new_keys = []
    for key in state_dict.keys():
        new_key = None
        if "gamma" in key:
            new_key = key.replace("gamma", "weight")
        if "beta" in key:
            new_key = key.replace("beta", "bias")
        if new_key:
            old_keys.append(key)
            new_keys.append(new_key)
    for old_key, new_key in zip(old_keys, new_keys):
        state_dict[new_key] = state_dict.pop(old_key)
714

715
716
717
718
719
720
721
722
723
    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
724
        set_module_kwargs = {}
725

726
        # We convert floating dtypes to the `dtype` passed. We want to keep the buffers/params
727
728
        # in int/uint/bool and not cast them.
        if dtype is not None and torch.is_floating_point(param):
729
730
            if (
                keep_in_fp32_modules is not None
731
732
733
                and any(
                    module_to_keep_in_fp32 in param_name.split(".") for module_to_keep_in_fp32 in keep_in_fp32_modules
                )
734
735
736
                and dtype == torch.float16
            ):
                param = param.to(torch.float32)
737
738
739
740
741

                # 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
742
743
            else:
                param = param.to(dtype)
744
745
746
747
748
749
750
751
752
753
754
755

        # For compatibility with PyTorch load_state_dict which converts state dict dtype to existing dtype in model
        if dtype is None:
            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:
                param = param.to(old_param.dtype)
756

757
758
        set_module_kwargs["value"] = param

759
760
761
762
763
764
765
766
767
768
769
        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]
770

771
        if param_device == "disk":
Sylvain Gugger's avatar
Sylvain Gugger committed
772
773
            if not is_safetensors:
                offload_index = offload_weight(param, param_name, offload_folder, offload_index)
774
775
        elif param_device == "cpu" and state_dict_index is not None:
            state_dict_index = offload_weight(param, param_name, state_dict_folder, state_dict_index)
776
        elif not is_quantized:
777
778
            # For backward compatibility with older versions of `accelerate`
            set_module_tensor_to_device(model, param_name, param_device, **set_module_kwargs)
779
        else:
780
781
782
783
784
785
            if param.dtype == torch.int8 and param_name.replace("weight", "SCB") in state_dict.keys():
                fp16_statistics = state_dict[param_name.replace("weight", "SCB")]
            else:
                fp16_statistics = None

            if "SCB" not in param_name:
786
                set_module_quantized_tensor_to_device(
787
788
                    model, param_name, param_device, value=param, fp16_statistics=fp16_statistics
                )
789
790

    return error_msgs, offload_index, state_dict_index
791
792


793
794
795
796
797
798
799
800
801
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


802
class ModuleUtilsMixin:
Julien Chaumond's avatar
Julien Chaumond committed
803
    """
804
    A few utilities for `torch.nn.Modules`, to be used as a mixin.
Julien Chaumond's avatar
Julien Chaumond committed
805
806
    """

807
808
809
810
    @staticmethod
    def _hook_rss_memory_pre_forward(module, *args, **kwargs):
        try:
            import psutil
Sylvain Gugger's avatar
Sylvain Gugger committed
811
        except ImportError:
812
813
814
815
816
817
818
819
820
821
822
            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
823
        except ImportError:
824
825
826
827
828
829
830
831
832
833
            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
834
835
836
        """
        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
837
838
        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()`.
839
840
841
842
843
844
845
        """
        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
846
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
847
        Reset the `mem_rss_diff` attribute of each module (see [`~modeling_utils.ModuleUtilsMixin.add_memory_hooks`]).
Sylvain Gugger's avatar
Sylvain Gugger committed
848
        """
849
850
851
852
853
        for module in self.modules():
            module.mem_rss_diff = 0
            module.mem_rss_post_forward = 0
            module.mem_rss_pre_forward = 0

854
    @property
Sylvain Gugger's avatar
Sylvain Gugger committed
855
    def device(self) -> torch.device:
856
        """
857
        `torch.device`: The device on which the module is (assuming that all the module parameters are on the same
858
        device).
859
        """
Lysandre Debut's avatar
Lysandre Debut committed
860
        return get_parameter_device(self)
861

862
    @property
863
    def dtype(self) -> torch.dtype:
864
        """
865
        `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).
866
        """
Lysandre Debut's avatar
Lysandre Debut committed
867
        return get_parameter_dtype(self)
868
869

    def invert_attention_mask(self, encoder_attention_mask: Tensor) -> Tensor:
Sylvain Gugger's avatar
Sylvain Gugger committed
870
871
872
873
        """
        Invert an attention mask (e.g., switches 0. and 1.).

        Args:
874
            encoder_attention_mask (`torch.Tensor`): An attention mask.
Sylvain Gugger's avatar
Sylvain Gugger committed
875
876

        Returns:
877
            `torch.Tensor`: The inverted attention mask.
Sylvain Gugger's avatar
Sylvain Gugger committed
878
        """
879
880
881
882
883
884
885
886
887
888
        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
889
        encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * torch.finfo(self.dtype).min
890

891
892
        return encoder_extended_attention_mask

893
    @staticmethod
894
895
896
897
898
899
900
    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
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
        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

921
    def get_extended_attention_mask(
922
        self, attention_mask: Tensor, input_shape: Tuple[int], device: torch.device = None, dtype: torch.float = None
923
    ) -> Tensor:
Sylvain Gugger's avatar
Sylvain Gugger committed
924
925
        """
        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
926
927

        Arguments:
928
            attention_mask (`torch.Tensor`):
Sylvain Gugger's avatar
Sylvain Gugger committed
929
                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
930
            input_shape (`Tuple[int]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
931
                The shape of the input to the model.
932
933

        Returns:
934
            `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
935
        """
Yih-Dar's avatar
Yih-Dar committed
936
937
938
        if dtype is None:
            dtype = self.dtype

939
940
941
942
943
944
        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
                )
945
946
947
948
949
950
951
952
953
        # 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:
954
                extended_attention_mask = ModuleUtilsMixin.create_extended_attention_mask_for_decoder(
955
956
                    input_shape, attention_mask, device
                )
957
958
959
960
            else:
                extended_attention_mask = attention_mask[:, None, None, :]
        else:
            raise ValueError(
961
                f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"
962
963
964
965
            )

        # 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
966
        # positions we want to attend and the dtype's smallest value for masked positions.
967
968
        # 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
969
970
        extended_attention_mask = extended_attention_mask.to(dtype=dtype)  # fp16 compatibility
        extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(dtype).min
971
972
        return extended_attention_mask

Sylvain Gugger's avatar
Sylvain Gugger committed
973
974
975
    def get_head_mask(
        self, head_mask: Optional[Tensor], num_hidden_layers: int, is_attention_chunked: bool = False
    ) -> Tensor:
976
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
977
978
979
        Prepare the head mask if needed.

        Args:
980
            head_mask (`torch.Tensor` with shape `[num_heads]` or `[num_hidden_layers x num_heads]`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
981
                The mask indicating if we should keep the heads or not (1.0 for keep, 0.0 for discard).
982
            num_hidden_layers (`int`):
Sylvain Gugger's avatar
Sylvain Gugger committed
983
                The number of hidden layers in the model.
984
            is_attention_chunked (`bool`, *optional*, defaults to `False`):
Sylvain Gugger's avatar
Sylvain Gugger committed
985
986
                Whether or not the attentions scores are computed by chunks or not.

987
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
988
989
            `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.
990
991
992
        """
        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
993
994
            if is_attention_chunked is True:
                head_mask = head_mask.unsqueeze(-1)
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
        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()}"
1008
        head_mask = head_mask.to(dtype=self.dtype)  # switch to float if need + fp16 compatibility
1009
1010
        return head_mask

1011
1012
1013
1014
1015
    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:
1016
            only_trainable (`bool`, *optional*, defaults to `False`):
1017
1018
                Whether or not to return only the number of trainable parameters

1019
            exclude_embeddings (`bool`, *optional*, defaults to `False`):
1020
1021
1022
                Whether or not to return only the number of non-embeddings parameters

        Returns:
1023
            `int`: The number of parameters.
1024
1025
        """

1026
1027
1028
1029
        if exclude_embeddings:
            embedding_param_names = [
                f"{name}.weight" for name, module_type in self.named_modules() if isinstance(module_type, nn.Embedding)
            ]
1030
            total_parameters = [
1031
1032
1033
                parameter for name, parameter in self.named_parameters() if name not in embedding_param_names
            ]
        else:
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
            total_parameters = list(self.parameters())

        total_numel = []
        is_loaded_in_4bit = getattr(self, "is_loaded_in_4bit", False)
        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"
1044
                    " make sure to install bitsandbytes with `pip install bitsandbytes`. You also need a GPU. "
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
                )

        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):
                    total_numel.append(param.numel() * 2)
                else:
                    total_numel.append(param.numel())

        return sum(total_numel)
1057
1058
1059
1060
1061
1062

    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:
1063
            inputs (`dict`): The model inputs.
1064
1065

        Returns:
1066
            `int`: The total number of tokens.
1067
        """
1068
1069
        if not hasattr(self, "warnings_issued"):
            self.warnings_issued = {}
1070
1071
        if self.main_input_name in input_dict:
            return input_dict[self.main_input_name].numel()
1072
        elif "estimate_tokens" not in self.warnings_issued:
1073
            logger.warning(
1074
1075
                "Could not estimate the number of tokens of the input, floating-point operations will not be computed"
            )
1076
1077
            self.warnings_issued["estimate_tokens"] = True
        return 0
1078
1079
1080
1081
1082
1083
1084

    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
1085
1086
        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
1087
        re-use e.g. Albert or Universal Transformers, or if doing long-range modeling with very high sequence lengths.
1088
1089

        Args:
1090
            batch_size (`int`):
1091
1092
                The batch size for the forward pass.

1093
            sequence_length (`int`):
1094
1095
                The number of tokens in each line of the batch.

1096
            exclude_embeddings (`bool`, *optional*, defaults to `True`):
1097
1098
1099
                Whether or not to count embedding and softmax operations.

        Returns:
1100
            `int`: The number of floating-point operations.
1101
1102
1103
1104
        """

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

Julien Chaumond's avatar
Julien Chaumond committed
1105

1106
class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin, PeftAdapterMixin):
1107
1108
    r"""
    Base class for all models.
1109

Sylvain Gugger's avatar
Sylvain Gugger committed
1110
1111
    [`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:
1112

1113
1114
        - resize the input embeddings,
        - prune heads in the self-attention heads.
1115

1116
    Class attributes (overridden by derived classes):
Sylvain Gugger's avatar
Sylvain Gugger committed
1117

Sylvain Gugger's avatar
Sylvain Gugger committed
1118
1119
1120
1121
        - **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:
1122

Sylvain Gugger's avatar
Sylvain Gugger committed
1123
1124
            - **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.
1125
            - **path** (`str`) -- A path to the TensorFlow checkpoint.
1126

Sylvain Gugger's avatar
Sylvain Gugger committed
1127
1128
        - **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.
1129
        - **is_parallelizable** (`bool`) -- A flag indicating whether this model supports model parallelization.
Sylvain Gugger's avatar
Sylvain Gugger committed
1130
1131
        - **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).
1132
    """
1133

1134
    config_class = None
1135
    base_model_prefix = ""
1136
    main_input_name = "input_ids"
1137
    _auto_class = None
1138
    _no_split_modules = None
1139
    _skip_keys_device_placement = None
1140
    _keep_in_fp32_modules = None
1141

1142
1143
    # 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.
1144
    _keys_to_ignore_on_load_missing = None
1145
1146
1147
    # 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.
1148
    _keys_to_ignore_on_load_unexpected = None
1149
1150
    # 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)
1151
    _keys_to_ignore_on_save = None
Sylvain Gugger's avatar
Sylvain Gugger committed
1152
1153
    # a list of `state_dict` keys that are potentially tied to another key in the state_dict.
    _tied_weights_keys = None
1154

1155
    is_parallelizable = False
1156
    supports_gradient_checkpointing = False
1157

1158
1159
1160
    # Flash Attention 2 support
    _supports_flash_attn_2 = False

1161
1162
1163
    # SDPA support
    _supports_sdpa = False

1164
1165
1166
    # Has support for a `Cache` instance as `past_key_values`
    _supports_cache_class = False

1167
    @property
1168
    def dummy_inputs(self) -> Dict[str, torch.Tensor]:
1169
        """
1170
        `Dict[str, torch.Tensor]`: Dummy inputs to do a forward pass in the network.
1171
        """
1172
        return {"input_ids": torch.tensor(DUMMY_INPUTS)}
1173

1174
1175
1176
1177
1178
1179
1180
    @property
    def framework(self) -> str:
        """
        :str: Identifies that this is a PyTorch model.
        """
        return "pt"

1181
    def __init__(self, config: PretrainedConfig, *inputs, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
1182
        super().__init__()
1183
1184
        if not isinstance(config, PretrainedConfig):
            raise ValueError(
1185
1186
1187
                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)`"
1188
            )
1189
        # Save config and origin of the pretrained weights if given in model
1190
1191
1192
        config = self._autoset_attn_implementation(
            config, torch_dtype=torch.get_default_dtype(), check_device_map=False
        )
1193
        self.config = config
1194

1195
        self.name_or_path = config.name_or_path
1196
        self.warnings_issued = {}
1197
        self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
1198
1199
1200
1201
        # 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)
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215

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

    def _backward_compatibility_gradient_checkpointing(self):
        if self.supports_gradient_checkpointing and getattr(self.config, "gradient_checkpointing", False):
            self.gradient_checkpointing_enable()
            # Remove the attribute now that is has been consumed, so it's no saved in the config.
            delattr(self.config, "gradient_checkpointing")
1216

1217
1218
1219
1220
1221
1222
    @classmethod
    def _from_config(cls, config, **kwargs):
        """
        All context managers that the model should be initialized under go here.

        Args:
1223
1224
            torch_dtype (`torch.dtype`, *optional*):
                Override the default `torch.dtype` and load the model under this dtype.
1225
1226
        """
        torch_dtype = kwargs.pop("torch_dtype", None)
1227
        use_flash_attention_2 = kwargs.pop("use_flash_attention_2", False)
1228
1229
1230
1231
1232
1233

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

1234
1235
1236
1237
1238
        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(
            config, use_flash_attention_2=use_flash_attention_2, check_device_map=False
        )
1239

1240
1241
1242
1243
1244
1245
        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
1246
            with deepspeed.zero.Init(config_dict_or_path=deepspeed_config()):
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
                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

1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
    @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)
1276
        requested_attn_implementation = None
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
        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.
1293
            requested_attn_implementation = config._attn_implementation_internal
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305

        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,
1306
                hard_check_only=False,
1307
1308
                check_device_map=check_device_map,
            )
1309
        elif requested_attn_implementation in [None, "sdpa"]:
1310
            # use_flash_attention_2 takes priority over SDPA, hence SDPA treated in this elif.
1311
1312
1313
1314
            config = cls._check_and_enable_sdpa(
                config, hard_check_only=False if requested_attn_implementation is None else True
            )
        else:
1315
1316
1317
1318
            config._attn_implementation = "eager"

        return config

1319
1320
1321
1322
1323
1324
1325
    @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:
1326
            dtype (`torch.dtype`):
1327
1328
1329
                a floating dtype to set to.

        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
1330
1331
            `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`.
1332

1333
1334
        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.
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
        """
        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

1346
    @property
1347
1348
    def base_model(self) -> nn.Module:
        """
1349
        `torch.nn.Module`: The main body of the model.
1350
        """
1351
        return getattr(self, self.base_model_prefix, self)
thomwolf's avatar
thomwolf committed
1352

1353
1354
    @classmethod
    def can_generate(cls) -> bool:
1355
1356
1357
1358
1359
1360
        """
        Returns whether this model can generate sequences with `.generate()`.

        Returns:
            `bool`: Whether this model can generate sequences with `.generate()`.
        """
1361
1362
1363
        # 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):
1364
1365
1366
            return False
        return True

1367
1368
    @classmethod
    def _check_and_enable_flash_attn_2(
1369
1370
1371
1372
1373
1374
        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,
1375
1376
    ) -> PretrainedConfig:
        """
1377
        Checks the availability of Flash Attention 2 and compatibility with the current model.
1378

1379
        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.
1380
1381
1382
        """
        if not cls._supports_flash_attn_2:
            raise ValueError(
1383
1384
1385
                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"
1386
1387
            )

1388
        if not is_flash_attn_2_available():
1389
1390
1391
            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."

1392
1393
1394
1395
1396
            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:
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
                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}")
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422

        _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:
            logger.warning(
                "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]:
1423
1424
1425
            logger.warning(
                "Flash Attention 2.0 only supports torch.float16 and torch.bfloat16 dtypes. "
                "No dtype was provided, you should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator."
1426
1427
            )

1428
1429
1430
        # 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":
1431
1432
            if torch.cuda.is_available():
                logger.warning(
1433
                    "You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU"
1434
1435
1436
1437
                    " after initializing it on CPU with `model.to('cuda')`."
                )
            else:
                raise ValueError(
1438
                    "You are attempting to use Flash Attention 2.0 with a model not initialized on GPU and with no GPU available. "
1439
1440
1441
1442
                    "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 (
1443
1444
            check_device_map
            and device_map is not None
1445
1446
1447
1448
1449
1450
1451
            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."
            )
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
        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(
                    f"{cls.__name__} does not support an attention implementation through torch.nn.functional.scaled_dot_product_attention yet. Please open an issue on GitHub to "
                    "request support for this architecture: https://github.com/huggingface/transformers/issues/new"
                )
            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"
1483
1484
        return config

1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
    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()

1502
    def get_input_embeddings(self) -> nn.Module:
1503
1504
1505
1506
        """
        Returns the model's input embeddings.

        Returns:
1507
            `nn.Module`: A torch module mapping vocabulary to hidden states.
thomwolf's avatar
thomwolf committed
1508
        """
1509
        base_model = getattr(self, self.base_model_prefix, self)
thomwolf's avatar
thomwolf committed
1510
1511
1512
1513
        if base_model is not self:
            return base_model.get_input_embeddings()
        else:
            raise NotImplementedError
thomwolf's avatar
thomwolf committed
1514

1515
    def set_input_embeddings(self, value: nn.Module):
1516
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
1517
        Set model's input embeddings.
1518
1519

        Args:
1520
            value (`nn.Module`): A module mapping vocabulary to hidden states.
thomwolf's avatar
thomwolf committed
1521
1522
1523
1524
1525
1526
        """
        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
1527

1528
    def get_output_embeddings(self) -> nn.Module:
1529
1530
1531
1532
        """
        Returns the model's output embeddings.

        Returns:
1533
            `nn.Module`: A torch module mapping hidden states to vocabulary.
thomwolf's avatar
thomwolf committed
1534
        """
1535
        return None  # Overwrite for models with output embeddings
thomwolf's avatar
thomwolf committed
1536

1537
1538
    def _init_weights(self, module):
        """
1539
1540
1541
1542
        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.
1543
        """
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
        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
1554

1555
    def tie_weights(self):
1556
1557
        """
        Tie the weights between the input embeddings and the output embeddings.
1558

Sylvain Gugger's avatar
Sylvain Gugger committed
1559
1560
        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
1561
        """
1562
1563
1564
1565
        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
1566

1567
        if getattr(self.config, "is_encoder_decoder", False) and getattr(self.config, "tie_encoder_decoder", False):
Weizhen's avatar
Weizhen committed
1568
1569
            if hasattr(self, self.base_model_prefix):
                self = getattr(self, self.base_model_prefix)
1570
1571
            self._tie_encoder_decoder_weights(self.encoder, self.decoder, self.base_model_prefix)

Sylvain Gugger's avatar
Sylvain Gugger committed
1572
1573
1574
1575
        for module in self.modules():
            if hasattr(module, "_tie_weights"):
                module._tie_weights()

1576
1577
1578
    @staticmethod
    def _tie_encoder_decoder_weights(encoder: nn.Module, decoder: nn.Module, base_model_prefix: str):
        uninitialized_encoder_weights: List[str] = []
Weizhen's avatar
Weizhen committed
1579
1580
        if decoder.__class__ != encoder.__class__:
            logger.info(
Sylvain Gugger's avatar
Sylvain Gugger committed
1581
1582
                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
1583
            )
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593

        def tie_encoder_to_decoder_recursively(
            decoder_pointer: nn.Module,
            encoder_pointer: nn.Module,
            module_name: str,
            uninitialized_encoder_weights: List[str],
            depth=0,
        ):
            assert isinstance(decoder_pointer, nn.Module) and isinstance(
                encoder_pointer, nn.Module
1594
            ), f"{decoder_pointer} and {encoder_pointer} have to be of type nn.Module"
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
            if hasattr(decoder_pointer, "weight"):
                assert hasattr(encoder_pointer, "weight")
                encoder_pointer.weight = decoder_pointer.weight
                if hasattr(decoder_pointer, "bias"):
                    assert hasattr(encoder_pointer, "bias")
                    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}"

1610
                all_encoder_weights = {module_name + "/" + sub_name for sub_name in encoder_modules.keys()}
1611
1612
1613
1614
1615
                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
1616
1617
1618
                        if not isinstance(decoder_modules[decoder_name], type(encoder_modules[encoder_name])) and len(
                            encoder_modules
                        ) != len(decoder_modules):
1619
1620
                            # 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
1621
                            # thus skip this step and subtract one layer pos from encoder
1622
1623
1624
1625
1626
1627
                            encoder_layer_pos -= 1
                            continue
                    elif name not in encoder_modules:
                        continue
                    elif depth > 500:
                        raise ValueError(
Sylvain Gugger's avatar
Sylvain Gugger committed
1628
1629
                            "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."
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
                        )
                    else:
                        decoder_name = encoder_name = name
                    tie_encoder_to_decoder_recursively(
                        decoder_modules[decoder_name],
                        encoder_modules[encoder_name],
                        module_name + "/" + name,
                        uninitialized_encoder_weights,
                        depth=depth + 1,
                    )
                    all_encoder_weights.remove(module_name + "/" + encoder_name)

                uninitialized_encoder_weights += list(all_encoder_weights)

        # tie weights recursively
        tie_encoder_to_decoder_recursively(decoder, encoder, base_model_prefix, uninitialized_encoder_weights)
        if len(uninitialized_encoder_weights) > 0:
            logger.warning(
                f"The following encoder weights were not tied to the decoder {uninitialized_encoder_weights}"
            )

1651
    def _tie_or_clone_weights(self, output_embeddings, input_embeddings):
Lysandre's avatar
Lysandre committed
1652
        """Tie or clone module weights depending of whether we are using TorchScript or not"""
thomwolf's avatar
thomwolf committed
1653
        if self.config.torchscript:
1654
            output_embeddings.weight = nn.Parameter(input_embeddings.weight.clone())
thomwolf's avatar
thomwolf committed
1655
        else:
1656
            output_embeddings.weight = input_embeddings.weight
thomwolf's avatar
thomwolf committed
1657

Sam Shleifer's avatar
Sam Shleifer committed
1658
        if getattr(output_embeddings, "bias", None) is not None:
1659
            output_embeddings.bias.data = nn.functional.pad(
1660
                output_embeddings.bias.data,
Lysandre's avatar
Lysandre committed
1661
1662
1663
1664
                (
                    0,
                    output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0],
                ),
1665
1666
                "constant",
                0,
1667
            )
1668
        if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
1669
            output_embeddings.out_features = input_embeddings.num_embeddings
1670

Marc Sun's avatar
Marc Sun committed
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
    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
        """
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
        _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
1698
1699
        return list(_no_split_modules)

1700
1701
1702
    def resize_token_embeddings(
        self, new_num_tokens: Optional[int] = None, pad_to_multiple_of: Optional[int] = None
    ) -> nn.Embedding:
1703
        """
1704
        Resizes input token embeddings matrix of the model if `new_num_tokens != config.vocab_size`.
1705

1706
        Takes care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
thomwolf's avatar
thomwolf committed
1707

1708
        Arguments:
1709
            new_num_tokens (`int`, *optional*):
1710
                The new number of tokens in the embedding matrix. Increasing the size will add newly initialized
Sylvain Gugger's avatar
Sylvain Gugger committed
1711
1712
                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.
1713
            pad_to_multiple_of (`int`, *optional*):
Arthur's avatar
Arthur committed
1714
1715
                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`.
1716
1717
1718
1719
1720

                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
1721
1722

        Return:
1723
            `torch.nn.Embedding`: Pointer to the input tokens Embeddings Module of the model.
thomwolf's avatar
thomwolf committed
1724
        """
1725
        model_embeds = self._resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
Arthur's avatar
Arthur committed
1726
        if new_num_tokens is None and pad_to_multiple_of is None:
thomwolf's avatar
thomwolf committed
1727
            return model_embeds
thomwolf's avatar
thomwolf committed
1728
1729

        # Update base model and current model config
Arthur's avatar
Arthur committed
1730
1731
        self.config.vocab_size = model_embeds.weight.shape[0]
        self.vocab_size = model_embeds.weight.shape[0]
thomwolf's avatar
thomwolf committed
1732
1733

        # Tie weights again if needed
1734
        self.tie_weights()
thomwolf's avatar
thomwolf committed
1735

thomwolf's avatar
thomwolf committed
1736
1737
        return model_embeds

1738
    def _resize_token_embeddings(self, new_num_tokens, pad_to_multiple_of=None):
thomwolf's avatar
thomwolf committed
1739
        old_embeddings = self.get_input_embeddings()
1740
        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens, pad_to_multiple_of)
1741
1742
1743
        if hasattr(old_embeddings, "_hf_hook"):
            hook = old_embeddings._hf_hook
            add_hook_to_module(new_embeddings, hook)
1744
1745
        old_embeddings_requires_grad = old_embeddings.weight.requires_grad
        new_embeddings.requires_grad_(old_embeddings_requires_grad)
thomwolf's avatar
thomwolf committed
1746
        self.set_input_embeddings(new_embeddings)
1747

1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
        # Update new_num_tokens with the actual size of new_embeddings
        if pad_to_multiple_of is not None:
            if is_deepspeed_zero3_enabled():
                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]

1758
1759
1760
        # 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()
1761
            new_lm_head = self._get_resized_lm_head(old_lm_head, new_num_tokens)
1762
1763
1764
            if hasattr(old_lm_head, "_hf_hook"):
                hook = old_lm_head._hf_hook
                add_hook_to_module(new_lm_head, hook)
1765
1766
            old_lm_head_requires_grad = old_lm_head.weight.requires_grad
            new_lm_head.requires_grad_(old_lm_head_requires_grad)
1767
1768
            self.set_output_embeddings(new_lm_head)

thomwolf's avatar
thomwolf committed
1769
        return self.get_input_embeddings()
1770

1771
    def _get_resized_embeddings(
1772
1773
1774
1775
        self,
        old_embeddings: nn.Embedding,
        new_num_tokens: Optional[int] = None,
        pad_to_multiple_of: Optional[int] = None,
1776
    ) -> nn.Embedding:
1777
1778
1779
        """
        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
1780
1781

        Args:
1782
            old_embeddings (`torch.nn.Embedding`):
1783
                Old embeddings to be resized.
1784
            new_num_tokens (`int`, *optional*):
1785
                New number of tokens in the embedding matrix.
1786
1787

                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
1788
                vectors from the end. If not provided or `None`, just returns a pointer to the input tokens
1789
                `torch.nn.Embedding` module of the model without doing anything.
1790
            pad_to_multiple_of (`int`, *optional*):
Arthur's avatar
Arthur committed
1791
1792
                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`.
1793
1794
1795
1796
1797
1798

                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

1799
1800

        Return:
1801
1802
            `torch.nn.Embedding`: Pointer to the resized Embedding Module or the old Embedding Module if
            `new_num_tokens` is `None`
1803
        """
1804
1805
1806
1807
1808
1809
1810
1811

        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]
1812
            new_num_tokens = ((new_num_tokens + pad_to_multiple_of - 1) // pad_to_multiple_of) * pad_to_multiple_of
1813
        else:
1814
            logger.info(
1815
                "You are resizing the embedding layer without providing a `pad_to_multiple_of` parameter. This means that the new embedding"
1816
                f" dimension will be {new_num_tokens}. This might induce some performance reduction as *Tensor Cores* will not be available."
1817
                " For more details about this, or help on choosing the correct value for resizing, refer to this guide:"
1818
1819
1820
                " https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html#requirements-tc"
            )

1821
1822
1823
        if new_num_tokens is None:
            return old_embeddings

1824
1825
1826
1827
1828
1829
1830
1831
        if is_deepspeed_zero3_enabled():
            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()

1832
        if old_num_tokens == new_num_tokens and not is_deepspeed_zero3_enabled():
1833
1834
            return old_embeddings

1835
1836
        if not isinstance(old_embeddings, nn.Embedding):
            raise TypeError(
Sylvain Gugger's avatar
Sylvain Gugger committed
1837
1838
1839
                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}."
1840
1841
            )

1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
        # 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

1860
1861
        # numbers of tokens to copy
        n = min(old_num_tokens, new_num_tokens)
1862

1863
1864
1865
        if is_deepspeed_zero3_enabled():
            import deepspeed

1866
1867
1868
            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, :]
1869
1870
        else:
            new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
1871
1872
1873

        return new_embeddings

1874
    def _get_resized_lm_head(
1875
1876
        self, old_lm_head: nn.Linear, new_num_tokens: Optional[int] = None, transposed: Optional[bool] = False
    ) -> nn.Linear:
1877
1878
1879
1880
1881
        """
        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:
1882
            old_lm_head (`torch.nn.Linear`):
1883
                Old lm head liner layer to be resized.
1884
            new_num_tokens (`int`, *optional*):
1885
1886
1887
                New number of tokens in the linear matrix.

                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
1888
                vectors from the end. If not provided or `None`, just returns a pointer to the input tokens
1889
1890
1891
                `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`.
1892
1893

        Return:
Sylvain Gugger's avatar
Sylvain Gugger committed
1894
1895
            `torch.nn.Linear`: Pointer to the resized Linear Module or the old Linear Module if `new_num_tokens` is
            `None`
1896
1897
1898
1899
        """
        if new_num_tokens is None:
            return old_lm_head

1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
        if is_deepspeed_zero3_enabled():
            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()
            )
1911

1912
        if old_num_tokens == new_num_tokens and not is_deepspeed_zero3_enabled():
1913
1914
1915
1916
            return old_lm_head

        if not isinstance(old_lm_head, nn.Linear):
            raise TypeError(
Sylvain Gugger's avatar
Sylvain Gugger committed
1917
1918
1919
                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}."
1920
1921
1922
1923
1924
1925
            )

        # 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

1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
        # 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)

1940
1941
        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)

1942
1943
1944
        if is_deepspeed_zero3_enabled():
            import deepspeed

1945
1946
            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):
1947
1948
1949
                self._copy_lm_head_original_to_resized(
                    new_lm_head, old_lm_head, num_tokens_to_copy, transposed, has_new_lm_head_bias
                )
1950
        else:
1951
1952
            self._copy_lm_head_original_to_resized(
                new_lm_head, old_lm_head, num_tokens_to_copy, transposed, has_new_lm_head_bias
1953
            )
1954
1955
1956

        return new_lm_head

1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
    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]

1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
    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`"
        )

1982
    def init_weights(self):
1983
        """
1984
1985
        If needed prunes and maybe initializes weights. If using a custom `PreTrainedModel`, you need to implement any
        initialization logic in `_init_weights`.
1986
        """
1987
1988
1989
1990
        # Prune heads if needed
        if self.config.pruned_heads:
            self.prune_heads(self.config.pruned_heads)

1991
1992
        if _init_weights:
            # Initialize weights
1993
            self.apply(self._initialize_weights)
1994
1995
1996
1997

            # Tie weights should be skipped when not initializing all weights
            # since from_pretrained(...) calls tie weights anyways
            self.tie_weights()
1998

1999
2000
2001
    def prune_heads(self, heads_to_prune: Dict[int, List[int]]):
        """
        Prunes heads of the base model.
2002

2003
        Arguments:
2004
            heads_to_prune (`Dict[int, List[int]]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2005
2006
2007
                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
2008
        """
2009
        # save new sets of pruned heads as union of previously stored pruned heads and newly pruned heads
2010
        for layer, heads in heads_to_prune.items():
2011
2012
2013
            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

2014
        self.base_model._prune_heads(heads_to_prune)
thomwolf's avatar
thomwolf committed
2015

2016
    def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):
2017
2018
2019
2020
2021
        """
        Activates gradient checkpointing for the current model.

        Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint
        activations".
2022
2023
2024
2025
2026
2027
2028

        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.
2029
2030
2031
        """
        if not self.supports_gradient_checkpointing:
            raise ValueError(f"{self.__class__.__name__} does not support gradient checkpointing.")
2032
2033
2034
2035

        if gradient_checkpointing_kwargs is None:
            gradient_checkpointing_kwargs = {}

2036
        gradient_checkpointing_func = functools.partial(checkpoint, **gradient_checkpointing_kwargs)
2037

2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
        # 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=True, gradient_checkpointing_func=gradient_checkpointing_func)
        else:
            self.apply(partial(self._set_gradient_checkpointing, value=True))
            logger.warn(
                "You are using an old version of the checkpointing format that is deprecated (We will also silently ignore `gradient_checkpointing_kwargs` in case you passed it)."
                "Please update to the new format on your modeling file. To use the new format, you need to completely remove the definition of the method `_set_gradient_checkpointing` in your model."
            )
2050

2051
2052
2053
2054
2055
2056
2057
        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()

2058
    def _set_gradient_checkpointing(self, enable: bool = True, gradient_checkpointing_func: Callable = checkpoint):
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
        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."
            )

2080
    def gradient_checkpointing_disable(self):
2081
2082
2083
2084
2085
2086
2087
        """
        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:
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
            # For old GC format (transformers < 4.35.0) for models that live on the Hub
            # we will fall back to the overwritten `_set_gradient_checkpointing` methid
            _is_using_old_format = "value" in inspect.signature(self._set_gradient_checkpointing).parameters
            if not _is_using_old_format:
                self._set_gradient_checkpointing(enable=False)
            else:
                logger.warn(
                    "You are using an old version of the checkpointing format that is deprecated (We will also silently ignore `gradient_checkpointing_kwargs` in case you passed it)."
                    "Please update to the new format on your modeling file. To use the new format, you need to completely remove the definition of the method `_set_gradient_checkpointing` in your model."
                )
                self.apply(partial(self._set_gradient_checkpointing, value=False))
2099

2100
2101
2102
        if getattr(self, "_hf_peft_config_loaded", False):
            self.disable_input_require_grads()

2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
    @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())

2113
2114
2115
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
2116
        is_main_process: bool = True,
2117
2118
        state_dict: Optional[dict] = None,
        save_function: Callable = torch.save,
Sylvain Gugger's avatar
Sylvain Gugger committed
2119
        push_to_hub: bool = False,
2120
        max_shard_size: Union[int, str] = "5GB",
2121
        safe_serialization: bool = True,
2122
        variant: Optional[str] = None,
2123
        token: Optional[Union[str, bool]] = None,
2124
        save_peft_format: bool = True,
Sylvain Gugger's avatar
Sylvain Gugger committed
2125
        **kwargs,
2126
    ):
2127
2128
        """
        Save a model and its configuration file to a directory, so that it can be re-loaded using the
2129
        [`~PreTrainedModel.from_pretrained`] class method.
2130

2131
        Arguments:
2132
            save_directory (`str` or `os.PathLike`):
2133
                Directory to which to save. Will be created if it doesn't exist.
2134
2135
2136
2137
            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.
2138
            state_dict (nested dictionary of `torch.Tensor`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2139
2140
2141
                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).
2142
            save_function (`Callable`):
2143
                The function to use to save the state dictionary. Useful on distributed training like TPUs when one
2144
2145
                need to replace `torch.save` by another method.
            push_to_hub (`bool`, *optional*, defaults to `False`):
2146
2147
2148
                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).
2149
            max_shard_size (`int` or `str`, *optional*, defaults to `"5GB"`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2150
2151
                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"`).
2152
2153
                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
2154
2155
2156
2157
2158
2159
2160
2161

                <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>

2162
            safe_serialization (`bool`, *optional*, defaults to `True`):
2163
                Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
2164
2165
            variant (`str`, *optional*):
                If specified, weights are saved in the format pytorch_model.<variant>.bin.
2166
2167
2168
            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`).
2169
2170
2171
2172
            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`.
2173
            kwargs (`Dict[str, Any]`, *optional*):
2174
                Additional key word arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
2175
        """
2176
2177
2178
2179
        use_auth_token = kwargs.pop("use_auth_token", None)

        if use_auth_token is not None:
            warnings.warn(
2180
2181
                "The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
                FutureWarning,
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
            )
            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
2192
2193
        _hf_peft_config_loaded = getattr(self, "_hf_peft_config_loaded", False)

2194
        # Checks if the model has been loaded in 8-bit
Younes Belkada's avatar
Younes Belkada committed
2195
2196
2197
2198
2199
2200
        if (
            getattr(self, "is_loaded_in_8bit", False)
            and not getattr(self, "is_8bit_serializable", False)
            and not _hf_peft_config_loaded
        ):
            raise ValueError(
2201
                "You are calling `save_pretrained` to a 8-bit converted model you may likely encounter unexepected"
Younes Belkada's avatar
Younes Belkada committed
2202
                " behaviors. If you want to save 8-bit models, make sure to have `bitsandbytes>0.37.2` installed."
2203
2204
            )

Younes Belkada's avatar
Younes Belkada committed
2205
2206
        # If the model has adapters attached, you can save the adapters
        if getattr(self, "is_loaded_in_4bit", False) and not _hf_peft_config_loaded:
2207
2208
2209
2210
            raise NotImplementedError(
                "You are calling `save_pretrained` on a 4-bit converted model. This is currently not supported"
            )

2211
2212
2213
        if getattr(self, "_awq_is_fused", False):
            raise ValueError("You cannot save an AWQ model that uses fused modules!")

2214
2215
2216
2217
2218
        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")
2219
2220
        if safe_serialization and not is_safetensors_available():
            raise ImportError("`safe_serialization` requires the `safetensors library: `pip install safetensors`.")
2221

2222
        if os.path.isfile(save_directory):
2223
            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
2224
            return
2225

2226
2227
        os.makedirs(save_directory, exist_ok=True)

2228
2229
        if push_to_hub:
            commit_message = kwargs.pop("commit_message", None)
2230
            repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
2231
            repo_id = self._create_repo(repo_id, **kwargs)
2232
            files_timestamps = self._get_files_timestamps(save_directory)
2233

Julien Chaumond's avatar
Julien Chaumond committed
2234
        # Only save the model itself if we are using distributed training
2235
        model_to_save = unwrap_model(self)
2236

2237
2238
2239
2240
2241
        # 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
2242
2243
2244
        # Attach architecture to the config
        model_to_save.config.architectures = [model_to_save.__class__.__name__]

2245
2246
2247
2248
2249
        # 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)

2250
        # Save the config
2251
        if is_main_process:
2252
2253
            if not _hf_peft_config_loaded:
                model_to_save.config.save_pretrained(save_directory)
2254
2255
            if self.can_generate():
                model_to_save.generation_config.save_pretrained(save_directory)
2256

2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
            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

2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
                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]
2282
2283
                current_peft_config.save_pretrained(save_directory)

2284
2285
2286
        # Save the model
        if state_dict is None:
            state_dict = model_to_save.state_dict()
2287

2288
2289
2290
2291
2292
        # 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)

2293
        # Handle the case where some state_dict keys shouldn't be saved
2294
        if self._keys_to_ignore_on_save is not None:
2295
            for ignore_key in self._keys_to_ignore_on_save:
2296
2297
                if ignore_key in state_dict.keys():
                    del state_dict[ignore_key]
2298
2299
2300
2301
2302
        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():
2303
2304
2305
2306
2307
2308
2309
                # 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)
2310
2311
2312
2313
2314
2315
2316

            # These are all the pointers of shared tensors.
            shared_ptrs = {ptr: names for ptr, names in ptrs.items() if len(names) > 1}
            warn_names = set()
            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.
Sylvain Gugger's avatar
Sylvain Gugger committed
2317
                if self._tied_weights_keys is not None:
2318
2319
                    found = 0
                    for name in sorted(names):
Sylvain Gugger's avatar
Sylvain Gugger committed
2320
                        matches_pattern = any(re.search(pat, name) for pat in self._tied_weights_keys)
2321
                        if matches_pattern and name in state_dict:
2322
2323
2324
                            found += 1
                            if found < len(names):
                                del state_dict[name]
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341

                # 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.
                found = 0
                for name in names:
                    if name in state_dict:
                        found += 1
                        if found > 1:
                            del state_dict[name]
                            warn_names.add(name)
            if len(warn_names) > 0:
                logger.warning_once(
                    f"Removed shared tensor {warn_names} while saving. This should be OK, but check by verifying that you don't receive any warning while reloading",
                )
2342

Sylvain Gugger's avatar
Sylvain Gugger committed
2343
        # Shard the model if it is too big.
2344
2345
2346
2347
2348
        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
2349

2350
        shards, index = shard_checkpoint(state_dict, max_shard_size=max_shard_size, weights_name=weights_name)
Sylvain Gugger's avatar
Sylvain Gugger committed
2351
2352
2353
2354

        # Clean the folder from a previous save
        for filename in os.listdir(save_directory):
            full_filename = os.path.join(save_directory, filename)
2355
2356
            # 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.
2357
            weights_no_suffix = weights_name.replace(".bin", "").replace(".safetensors", "")
2358
2359
2360

            # 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", "")
2361
            reg = re.compile(r"(.*?)-\d{5}-of-\d{5}")
2362

2363
            if (
2364
                filename.startswith(weights_no_suffix)
2365
2366
2367
                and os.path.isfile(full_filename)
                and filename not in shards.keys()
                and is_main_process
2368
                and reg.fullmatch(filename_no_suffix) is not None
2369
            ):
Sylvain Gugger's avatar
Sylvain Gugger committed
2370
                os.remove(full_filename)
2371

Sylvain Gugger's avatar
Sylvain Gugger committed
2372
2373
        # Save the model
        for shard_file, shard in shards.items():
2374
2375
2376
2377
2378
2379
            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
2380
2381

        if index is None:
2382
2383
            path_to_weights = os.path.join(save_directory, _add_variant(WEIGHTS_NAME, variant))
            logger.info(f"Model weights saved in {path_to_weights}")
Sylvain Gugger's avatar
Sylvain Gugger committed
2384
        else:
2385
            save_index_file = SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
2386
            save_index_file = os.path.join(save_directory, _add_variant(save_index_file, variant))
Sylvain Gugger's avatar
Sylvain Gugger committed
2387
2388
2389
2390
2391
2392
2393
2394
2395
            # Save the index as well
            with open(save_index_file, "w", encoding="utf-8") as f:
                content = json.dumps(index, indent=2, sort_keys=True) + "\n"
                f.write(content)
            logger.info(
                f"The model is bigger than the maximum size per checkpoint ({max_shard_size}) and is going to be "
                f"split in {len(shards)} checkpoint shards. You can find where each parameters has been saved in the "
                f"index located at {save_index_file}."
            )
2396

Sylvain Gugger's avatar
Sylvain Gugger committed
2397
        if push_to_hub:
2398
            self._upload_modified_files(
2399
2400
2401
2402
                save_directory,
                repo_id,
                files_timestamps,
                commit_message=commit_message,
2403
                token=token,
2404
            )
Sylvain Gugger's avatar
Sylvain Gugger committed
2405

2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
    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

2424
    @wraps(torch.nn.Module.cuda)
2425
2426
    def cuda(self, *args, **kwargs):
        # Checks if the model has been loaded in 8-bit
Marc Sun's avatar
Marc Sun committed
2427
        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
2428
2429
2430
2431
2432
2433
2434
            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)

2435
    @wraps(torch.nn.Module.to)
2436
2437
    def to(self, *args, **kwargs):
        # Checks if the model has been loaded in 8-bit
Marc Sun's avatar
Marc Sun committed
2438
        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
2439
            raise ValueError(
Marc Sun's avatar
Marc Sun committed
2440
                "`.to` is not supported for `4-bit` or `8-bit` bitsandbytes models. Please use the model as it is, since the"
2441
2442
                " model has already been set to the correct devices and casted to the correct `dtype`."
            )
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
        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)
2462
2463

    def half(self, *args):
Marc Sun's avatar
Marc Sun committed
2464
        # Checks if the model is quantized
2465
        if getattr(self, "is_quantized", False):
2466
            raise ValueError(
Marc Sun's avatar
Marc Sun committed
2467
                "`.half()` is not supported for quantized model. Please use the model as it is, since the"
2468
2469
2470
2471
2472
2473
                " 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
2474
        # Checks if the model is quantized
2475
        if getattr(self, "is_quantized", False):
2476
            raise ValueError(
Marc Sun's avatar
Marc Sun committed
2477
                "`.float()` is not supported for quantized model. Please use the model as it is, since the"
2478
2479
2480
2481
2482
                " model has already been casted to the correct `dtype`."
            )
        else:
            return super().float(*args)

2483
    @classmethod
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
    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,
    ):
2498
2499
        r"""
        Instantiate a pretrained pytorch model from a pre-trained model configuration.
2500

Sylvain Gugger's avatar
Sylvain Gugger committed
2501
2502
        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()`.
2503

2504
        The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
2505
2506
        pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
        task.
2507

2508
        The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
2509
        weights are discarded.
2510

2511
        Parameters:
2512
            pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
2513
2514
                Can be either:

2515
                    - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
Sylvain Gugger's avatar
Sylvain Gugger committed
2516
2517
                      Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
                      user or organization name, like `dbmdz/bert-base-german-cased`.
2518
2519
2520
                    - 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
2521
2522
2523
                      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.
2524
                    - 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
2525
2526
                      `./flax_model/` containing `flax_model.msgpack`). In this case, `from_flax` should be set to
                      `True`.
2527
2528
2529
2530
2531
                    - `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*):
2532
2533
                Can be either:

2534
2535
                    - an instance of a class derived from [`PretrainedConfig`],
                    - a string or path valid as input to [`~PretrainedConfig.from_pretrained`].
2536

2537
                Configuration for the model to use instead of an automatically loaded configuration. Configuration can
2538
2539
                be automatically loaded when:

2540
                    - The model is a model provided by the library (loaded with the *model id* string of a pretrained
2541
                      model).
Sylvain Gugger's avatar
Sylvain Gugger committed
2542
2543
                    - The model was saved using [`~PreTrainedModel.save_pretrained`] and is reloaded by supplying the
                      save directory.
2544
2545
2546
                    - 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*):
2547
2548
2549
                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
2550
                weights. In this case though, you should check if using [`~PreTrainedModel.save_pretrained`] and
2551
2552
                [`~PreTrainedModel.from_pretrained`] is not a simpler option.
            cache_dir (`Union[str, os.PathLike]`, *optional*):
2553
2554
                Path to a directory in which a downloaded pretrained model configuration should be cached if the
                standard cache should not be used.
2555
            from_tf (`bool`, *optional*, defaults to `False`):
2556
                Load the model weights from a TensorFlow checkpoint save file (see docstring of
2557
2558
                `pretrained_model_name_or_path` argument).
            from_flax (`bool`, *optional*, defaults to `False`):
2559
                Load the model weights from a Flax checkpoint save file (see docstring of
2560
2561
                `pretrained_model_name_or_path` argument).
            ignore_mismatched_sizes (`bool`, *optional*, defaults to `False`):
2562
2563
2564
                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).
2565
            force_download (`bool`, *optional*, defaults to `False`):
2566
2567
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
2568
            resume_download (`bool`, *optional*, defaults to `False`):
2569
2570
                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
                file exists.
2571
            proxies (`Dict[str, str]`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
2572
2573
                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.
2574
            output_loading_info(`bool`, *optional*, defaults to `False`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2575
                Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
2576
            local_files_only(`bool`, *optional*, defaults to `False`):
Stas Bekman's avatar
Stas Bekman committed
2577
                Whether or not to only look at local files (i.e., do not try to download the model).
2578
            token (`str` or `bool`, *optional*):
2579
2580
                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`).
2581
            revision (`str`, *optional*, defaults to `"main"`):
Julien Chaumond's avatar
Julien Chaumond committed
2582
                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
2583
                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
Julien Chaumond's avatar
Julien Chaumond committed
2584
                identifier allowed by git.
2585
2586
2587
2588
2589
2590
2591

                <Tip>

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

                </Tip>

2592
            mirror (`str`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
2593
2594
2595
                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.
2596
            _fast_init(`bool`, *optional*, defaults to `True`):
2597
2598
                Whether or not to disable fast initialization.

2599
2600
                <Tip warning={true}>

Sylvain Gugger's avatar
Sylvain Gugger committed
2601
2602
2603
                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.
2604

2605
                </Tip>
2606

2607
2608
2609
            > Parameters for big model inference

            low_cpu_mem_usage(`bool`, *optional*):
2610
2611
2612
                Tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
                This is an experimental feature and a subject to change at any moment.
            torch_dtype (`str` or `torch.dtype`, *optional*):
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
                Override the default `torch.dtype` and load the model under a specific `dtype`. The different options
                are:

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

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

                <Tip>

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

                </Tip>

2634
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]` or `int` or `torch.device`, *optional*):
2635
2636
                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
2637
2638
2639
                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.
2640

2641
2642
                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
2643
                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
2644
2645
2646
            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.
2647
2648
            offload_folder (`str` or `os.PathLike`, *optional*):
                If the `device_map` contains any value `"disk"`, the folder where we will offload weights.
2649
            offload_state_dict (`bool`, *optional*):
2650
                If `True`, will temporarily offload the CPU state dict to the hard drive to avoid getting out of CPU
2651
2652
                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.
2653
2654
            load_in_8bit (`bool`, *optional*, defaults to `False`):
                If `True`, will convert the loaded model into mixed-8bit quantized model. To use this feature please
2655
2656
2657
2658
                install `bitsandbytes` (`pip install -U bitsandbytes`).
            load_in_4bit (`bool`, *optional*, defaults to `False`):
                If `True`, will convert the loaded model into 4bit precision quantized model. To use this feature
                install the latest version of `bitsandbytes` (`pip install -U bitsandbytes`).
Marc Sun's avatar
Marc Sun committed
2659
2660
2661
            quantization_config (`Union[QuantizationConfigMixin,Dict]`, *optional*):
                A dictionary of configuration parameters or a QuantizationConfigMixin object for quantization (e.g
                bitsandbytes, gptq)
2662
2663
2664
            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.
2665
2666
2667
            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`.
2668
2669
2670
            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`.
2671

2672
            kwargs (remaining dictionary of keyword arguments, *optional*):
2673
                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
2674
                `output_attentions=True`). Behaves differently depending on whether a `config` is provided or
2675
2676
                automatically loaded:

2677
2678
                    - 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
2679
                      already been done)
2680
                    - If a configuration is not provided, `kwargs` will be first passed to the configuration class
Sylvain Gugger's avatar
Sylvain Gugger committed
2681
2682
2683
2684
                      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.
2685
2686
2687

        <Tip>

Sylvain Gugger's avatar
Sylvain Gugger committed
2688
2689
        Activate the special ["offline-mode"](https://huggingface.co/transformers/installation.html#offline-mode) to
        use this method in a firewalled environment.
2690
2691
2692
2693
2694
2695
2696

        </Tip>

        Examples:

        ```python
        >>> from transformers import BertConfig, BertModel
Sylvain Gugger's avatar
Sylvain Gugger committed
2697

2698
        >>> # Download model and configuration from huggingface.co and cache.
Sylvain Gugger's avatar
Sylvain Gugger committed
2699
        >>> model = BertModel.from_pretrained("bert-base-uncased")
2700
        >>> # Model was saved using *save_pretrained('./test/saved_model/')* (for example purposes, not runnable).
Sylvain Gugger's avatar
Sylvain Gugger committed
2701
        >>> model = BertModel.from_pretrained("./test/saved_model/")
2702
        >>> # Update configuration during loading.
Sylvain Gugger's avatar
Sylvain Gugger committed
2703
        >>> model = BertModel.from_pretrained("bert-base-uncased", output_attentions=True)
2704
2705
        >>> 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
2706
2707
        >>> 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)
2708
        >>> # Loading from a Flax checkpoint file instead of a PyTorch model (slower)
Sylvain Gugger's avatar
Sylvain Gugger committed
2709
        >>> model = BertModel.from_pretrained("bert-base-uncased", from_flax=True)
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
        ```

        * `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

        """
2728
2729
        state_dict = kwargs.pop("state_dict", None)
        from_tf = kwargs.pop("from_tf", False)
2730
        from_flax = kwargs.pop("from_flax", False)
2731
2732
2733
        resume_download = kwargs.pop("resume_download", False)
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
2734
        use_auth_token = kwargs.pop("use_auth_token", None)
2735
        trust_remote_code = kwargs.pop("trust_remote_code", None)
Sylvain Gugger's avatar
Sylvain Gugger committed
2736
        _ = kwargs.pop("mirror", None)
2737
2738
        from_pipeline = kwargs.pop("_from_pipeline", None)
        from_auto_class = kwargs.pop("_from_auto", False)
2739
        _fast_init = kwargs.pop("_fast_init", True)
2740
        torch_dtype = kwargs.pop("torch_dtype", None)
2741
2742
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", None)
        device_map = kwargs.pop("device_map", None)
2743
        max_memory = kwargs.pop("max_memory", None)
2744
        offload_folder = kwargs.pop("offload_folder", None)
2745
2746
        offload_state_dict = kwargs.pop("offload_state_dict", False)
        load_in_8bit = kwargs.pop("load_in_8bit", False)
2747
        load_in_4bit = kwargs.pop("load_in_4bit", False)
2748
        quantization_config = kwargs.pop("quantization_config", None)
2749
        subfolder = kwargs.pop("subfolder", "")
2750
        commit_hash = kwargs.pop("_commit_hash", None)
2751
        variant = kwargs.pop("variant", None)
2752
        adapter_kwargs = kwargs.pop("adapter_kwargs", {})
2753
        adapter_name = kwargs.pop("adapter_name", "default")
2754
        use_flash_attention_2 = kwargs.pop("use_flash_attention_2", False)
2755

2756
2757
2758
        if is_fsdp_enabled():
            low_cpu_mem_usage = True

2759
2760
        if use_auth_token is not None:
            warnings.warn(
2761
2762
                "The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",
                FutureWarning,
2763
2764
2765
2766
2767
2768
2769
            )
            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

2770
        if token is not None and adapter_kwargs is not None and "token" not in adapter_kwargs:
2771
2772
            adapter_kwargs["token"] = token

2773
2774
        if use_safetensors is None and not is_safetensors_available():
            use_safetensors = False
2775

2776
        if is_bitsandbytes_available():
2777
            is_8bit_serializable = version.parse(importlib.metadata.version("bitsandbytes")) > version.parse("0.37.2")
2778
2779
2780
        else:
            is_8bit_serializable = False

2781
2782
2783
2784
2785
        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."
            )
2786

2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
        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,
                    _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)

2808
        if is_peft_available():
2809
2810
            _adapter_model_path = adapter_kwargs.pop("_adapter_model_path", None)

2811
2812
2813
2814
2815
2816
2817
2818
2819
            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,
2820
                    **adapter_kwargs,
2821
2822
                )
            if _adapter_model_path is not None and os.path.isfile(_adapter_model_path):
2823
                with open(_adapter_model_path, "r", encoding="utf-8") as f:
2824
                    _adapter_model_path = pretrained_model_name_or_path
2825
                    pretrained_model_name_or_path = json.load(f)["base_model_name_or_path"]
2826
2827
        else:
            _adapter_model_path = None
2828

2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
        # 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}

2848
2849
2850
2851
2852
2853
2854
        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:
2855
            if device_map is not None:
2856
2857
                # The max memory utils require PyTorch >= 1.10 to have torch.cuda.mem_get_info.
                require_version_core("torch>=1.10")
2858
2859
2860
2861
2862
2863
2864
2865
2866

            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`"
                )
2867

Marc Sun's avatar
Marc Sun committed
2868
        quantization_method_from_args = None
2869

Marc Sun's avatar
Marc Sun committed
2870
2871
2872
2873
2874
2875
2876
        if quantization_config is not None:
            quantization_method_from_args = getattr(
                quantization_config, "quant_method", QuantizationMethod.BITS_AND_BYTES
            )

        if quantization_config is None and (load_in_8bit or load_in_4bit):
            quantization_method_from_args = QuantizationMethod.BITS_AND_BYTES
2877
            quantization_config, kwargs = BitsAndBytesConfig.from_dict(
2878
2879
2880
                config_dict={"load_in_8bit": load_in_8bit, "load_in_4bit": load_in_4bit},
                return_unused_kwargs=True,
                **kwargs,
2881
            )
Marc Sun's avatar
Marc Sun committed
2882
        elif quantization_method_from_args == QuantizationMethod.BITS_AND_BYTES:
2883
            load_in_8bit = quantization_config.load_in_8bit
2884
            load_in_4bit = quantization_config.load_in_4bit
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895

            quantization_config_kwargs = {
                k: v for k, v in kwargs.items() if k in inspect.signature(BitsAndBytesConfig).parameters
            }

            if len(quantization_config_kwargs) > 0:
                raise ValueError(
                    "You can't pass `load_in_8bit` or any other `BitsAndBytesConfig` argument as a kwarg when passing "
                    "`quantization_config` argument at the same time."
                )

2896
        if load_in_8bit or load_in_4bit:
2897
2898
            if not torch.cuda.is_available():
                raise RuntimeError("No GPU found. A GPU is needed for quantization.")
2899
2900
2901
2902
            if not (is_accelerate_available() and is_bitsandbytes_available()):
                raise ImportError(
                    "Using `load_in_8bit=True` requires Accelerate: `pip install accelerate` and the latest version of"
                    " bitsandbytes `pip install -i https://test.pypi.org/simple/ bitsandbytes` or"
2903
                    " `pip install bitsandbytes`."
2904
                )
2905
2906

            if torch_dtype is None:
2907
                # We force the `dtype` to be float16, this is a requirement from `bitsandbytes`
2908
                logger.info(
2909
                    f"Overriding torch_dtype={torch_dtype} with `torch_dtype=torch.float16` due to "
2910
                    "requirements of `bitsandbytes` to enable model loading in 8-bit or 4-bit. "
2911
2912
                    "Pass your own torch_dtype to specify the dtype of the remaining non-linear layers or pass"
                    " torch_dtype=torch.float16 to remove this warning."
2913
                )
2914
                torch_dtype = torch.float16
2915

2916
            if device_map is None:
2917
                device_map = {"": torch.cuda.current_device()}
2918
                logger.info(
2919
2920
                    "The device_map was not initialized. "
                    "Setting device_map to {'':torch.cuda.current_device()}. "
2921
                    "If you want to use the model for inference, please set device_map ='auto' "
2922
                )
2923
2924
2925
                if low_cpu_mem_usage is None:
                    low_cpu_mem_usage = True

2926
2927
            if from_tf or from_flax:
                raise ValueError(
2928
                    "Converting into 4-bit or 8-bit weights from tf/flax weights is currently not supported, please make"
2929
2930
2931
                    " sure the weights are in PyTorch format."
                )

2932
2933
2934
        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
2935

2936
2937
2938
2939
        if is_offline_mode() and not local_files_only:
            logger.info("Offline mode: forcing local_files_only=True")
            local_files_only = True

2940
2941
2942
        # 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
2943
            config, model_kwargs = cls.config_class.from_pretrained(
2944
2945
2946
                config_path,
                cache_dir=cache_dir,
                return_unused_kwargs=True,
2947
                force_download=force_download,
2948
                resume_download=resume_download,
2949
                proxies=proxies,
2950
                local_files_only=local_files_only,
2951
                token=token,
Julien Chaumond's avatar
Julien Chaumond committed
2952
                revision=revision,
2953
                subfolder=subfolder,
2954
2955
                _from_auto=from_auto_class,
                _from_pipeline=from_pipeline,
2956
                **kwargs,
2957
2958
            )
        else:
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
            # In case one passes a config to `from_pretrained` + "attn_implementation"
            # override the `_attn_implementation` attribute to `attn_implementation` of the kwargs
            # Please see: https://github.com/huggingface/transformers/issues/28038

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

            kwarg_attn_imp = kwargs.pop("attn_implementation", None)
            if kwarg_attn_imp is not None and config._attn_implementation != kwarg_attn_imp:
                config._attn_implementation = kwarg_attn_imp
2971
            model_kwargs = kwargs
2972

Marc Sun's avatar
Marc Sun committed
2973
2974
2975
2976
2977
2978
        quantizer = None
        quantization_method_from_config = None
        if hasattr(config, "quantization_config"):
            quantization_method_from_config = config.quantization_config.get(
                "quant_method", QuantizationMethod.BITS_AND_BYTES
            )
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989

        if (
            quantization_method_from_args is not None
            and quantization_method_from_args == QuantizationMethod.AWQ
            and quantization_method_from_config is None
        ):
            raise ValueError(
                "You cannot quantize with AWQ a non-quantized model using transformers, please refer to the quantization documentation"
                " to read more about how to quantize models with AWQ algorithm https://huggingface.co/docs/transformers/main_classes/quantization"
            )

2990
2991
2992
2993
2994
2995
        if quantization_method_from_config is not None and quantization_method_from_args is not None:
            if quantization_method_from_config != quantization_method_from_args:
                raise ValueError(
                    f"The model is already quantized with {quantization_method_from_config}. "
                    f"You can't quantize it again with {quantization_method_from_args}"
                )
2996
2997
2998
2999
3000

        if (
            quantization_method_from_config in (QuantizationMethod.GPTQ, QuantizationMethod.AWQ)
            and quantization_method_from_args is not None
        ):
Marc Sun's avatar
Marc Sun committed
3001
3002
3003
3004
3005
            loading_attr_dict = quantization_config.get_loading_attributes()
            for attr, val in loading_attr_dict.items():
                config.quantization_config[attr] = val
            quantization_method_from_args = None
            logger.warning(
3006
3007
3008
                f"You passed `quantization_config` to `from_pretrained` but the model you're loading already has a "
                f"`quantization_config` attribute and has already quantized weights. However, loading attributes"
                f" (e.g. {list(loading_attr_dict.keys())}) will be overwritten with the one you passed to `from_pretrained`. The rest will be ignored."
Marc Sun's avatar
Marc Sun committed
3009
3010
3011
3012
3013
            )
        if (
            quantization_method_from_args == QuantizationMethod.GPTQ
            or quantization_method_from_config == QuantizationMethod.GPTQ
        ):
3014
3015
            gptq_supports_cpu = version.parse(importlib.metadata.version("auto-gptq")) > version.parse("0.4.2")
            if not gptq_supports_cpu and not torch.cuda.is_available():
Marc Sun's avatar
Marc Sun committed
3016
3017
3018
                raise RuntimeError("GPU is required to quantize or run quantize model.")
            elif not (is_optimum_available() and is_auto_gptq_available()):
                raise ImportError(
3019
3020
3021
3022
3023
                    "Loading a GPTQ quantized model requires optimum (`pip install optimum`) and auto-gptq library (`pip install auto-gptq`)"
                )
            elif version.parse(importlib.metadata.version("auto_gptq")) < version.parse("0.4.2"):
                raise ImportError(
                    "You need a version of auto_gptq >= 0.4.2 to use GPTQ: `pip install --upgrade auto-gptq`"
Marc Sun's avatar
Marc Sun committed
3024
3025
3026
3027
3028
3029
3030
                )
            else:
                # Need to protect the import
                from optimum.gptq import GPTQQuantizer
            if quantization_method_from_config == QuantizationMethod.GPTQ:
                quantization_config = GPTQConfig.from_dict(config.quantization_config)
                config.quantization_config = quantization_config
3031
3032
3033
3034
            if torch_dtype is None:
                torch_dtype = torch.float16
            else:
                logger.info("We suggest you to set `torch_dtype=torch.float16` for better efficiency with GPTQ.")
Marc Sun's avatar
Marc Sun committed
3035
            quantizer = GPTQQuantizer.from_dict(quantization_config.to_dict_optimum())
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
        elif quantization_method_from_config == QuantizationMethod.AWQ:
            if not torch.cuda.is_available():
                raise RuntimeError("GPU is required to run AWQ quantized model.")

            if not is_auto_awq_available():
                raise ImportError("Loading an AWQ quantized model requires auto-awq library (`pip install autoawq`)")

            if not is_accelerate_available():
                raise ImportError("Loading an AWQ quantized model requires accelerate (`pip install accelerate`)")

            if device_map is None:
                logger.warning(
                    "You have loaded an AWQ model on CPU and have a CUDA device available, make sure to set "
                    "your model on a GPU device in order to run your model."
                )
            elif device_map is not None:
                if isinstance(device_map, dict) and ("cpu" in device_map.values() or "disk" in device_map.values()):
                    raise ValueError(
                        "You are attempting to load an AWQ model with a device_map that contains a CPU or disk device."
                        " This is not supported. Please remove the CPU or disk device from the device_map."
                    )

            if torch_dtype is None:
                torch_dtype = torch.float16
            else:
                logger.info("We suggest you to set `torch_dtype=torch.float16` for better efficiency with AWQ.")

            # Force-set to `True` for more mem efficiency
            if low_cpu_mem_usage is None:
                low_cpu_mem_usage = True
Marc Sun's avatar
Marc Sun committed
3066
3067
3068
3069
3070
3071
3072

        if (
            is_8bit_serializable
            and quantization_method_from_args == QuantizationMethod.BITS_AND_BYTES
            and load_in_8bit
        ):
            if quantization_method_from_config == QuantizationMethod.BITS_AND_BYTES:
3073
3074
3075
3076
3077
3078
                logger.warning(
                    "You passed `quantization_config` to `from_pretrained` but the model you're loading already has a"
                    " `quantization_config` attribute. The `quantization_config` attribute will be overwritten with the"
                    " one you passed to `from_pretrained`."
                )
            config.quantization_config = quantization_config
Marc Sun's avatar
Marc Sun committed
3079
3080
3081
3082
3083
        elif (
            is_8bit_serializable
            and not load_in_8bit
            and quantization_method_from_config == QuantizationMethod.BITS_AND_BYTES
        ):
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
            quantization_config = config.quantization_config
            if isinstance(quantization_config, dict):
                quantization_config = BitsAndBytesConfig.from_dict(quantization_config, return_unused_kwargs=False)
            elif isinstance(quantization_config, BitsAndBytesConfig):
                pass
            else:
                raise ValueError(
                    f"Invalid type for `quantization_config`: {type(quantization_config)}. Should be a `dict` or a"
                    " `BitsAndBytesConfig` instance."
                )

            load_in_8bit = quantization_config.load_in_8bit

            if load_in_8bit:
3098
3099
                if torch_dtype is None:
                    torch_dtype = torch.float16
3100
                if device_map is None:
3101
3102
3103
3104
3105
                    if torch.cuda.is_available():
                        device_map = {"": torch.cuda.current_device()}
                    else:
                        raise RuntimeError("No GPU found. A GPU is needed for quantization.")
                    logger.info(
3106
3107
                        "The device_map was not initialized. "
                        "Setting device_map to {'':torch.cuda.current_device()}. "
3108
3109
3110
3111
                        "If you want to use the model for inference, please set device_map ='auto' "
                    )
                    if low_cpu_mem_usage is None:
                        low_cpu_mem_usage = True
3112

Marc Sun's avatar
Marc Sun committed
3113
3114
3115
3116
3117
        elif (
            not is_8bit_serializable
            and not load_in_8bit
            and quantization_method_from_config == QuantizationMethod.BITS_AND_BYTES
        ):
3118
3119
3120
3121
3122
3123
            logger.warning(
                "Detected the presence of a `quantization_config` attribute in the model's configuration but you don't have the correct"
                " `bitsandbytes` version to support int8 serialization. Please install the latest version of `bitsandbytes` with "
                " `pip install --upgrade bitsandbytes`."
            )

Sylvain Gugger's avatar
Sylvain Gugger committed
3124
3125
3126
3127
        # 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
3128
        # Load model
Yih-Dar's avatar
Yih-Dar committed
3129
3130
        loading_info = None

3131
3132
3133
3134
        # Keep in fp32 modules
        keep_in_fp32_modules = None
        use_keep_in_fp32_modules = False

thomwolf's avatar
thomwolf committed
3135
        if pretrained_model_name_or_path is not None:
3136
            pretrained_model_name_or_path = str(pretrained_model_name_or_path)
3137
3138
            is_local = os.path.isdir(pretrained_model_name_or_path)
            if is_local:
3139
3140
3141
                if from_tf and os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, TF_WEIGHTS_NAME + ".index")
                ):
3142
                    # Load from a TF 1.0 checkpoint in priority if from_tf
3143
3144
3145
3146
                    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)
                ):
3147
                    # Load from a TF 2.0 checkpoint in priority if from_tf
3148
3149
3150
3151
                    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)
                ):
3152
                    # Load from a Flax checkpoint in priority if from_flax
3153
                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)
3154
                elif use_safetensors is not False and os.path.isfile(
3155
                    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant))
3156
3157
                ):
                    # Load from a safetensors checkpoint
3158
3159
3160
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant)
                    )
3161
                elif use_safetensors is not False and os.path.isfile(
3162
3163
3164
                    os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
                    )
3165
3166
                ):
                    # Load from a sharded safetensors checkpoint
3167
3168
3169
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
                    )
3170
                    is_sharded = True
3171
3172
3173
                elif os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_NAME, variant))
                ):
thomwolf's avatar
thomwolf committed
3174
                    # Load from a PyTorch checkpoint
3175
3176
3177
3178
3179
3180
                    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
3181
                    # Load from a sharded PyTorch checkpoint
3182
3183
3184
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_INDEX_NAME, variant)
                    )
Sylvain Gugger's avatar
Sylvain Gugger committed
3185
                    is_sharded = True
3186
3187
                # At this stage we don't have a weight file so we will raise an error.
                elif os.path.isfile(
3188
3189
                    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)):
3190
                    raise EnvironmentError(
3191
3192
3193
                        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."
3194
                    )
3195
                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)):
3196
                    raise EnvironmentError(
3197
3198
3199
                        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."
3200
                    )
3201
3202
3203
3204
3205
                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
3206
                else:
3207
                    raise EnvironmentError(
3208
3209
3210
                        f"Error no file named {_add_variant(WEIGHTS_NAME, variant)}, {TF2_WEIGHTS_NAME},"
                        f" {TF_WEIGHTS_NAME + '.index'} or {FLAX_WEIGHTS_NAME} found in directory"
                        f" {pretrained_model_name_or_path}."
3211
                    )
3212
            elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path)):
3213
                archive_file = pretrained_model_name_or_path
3214
                is_local = True
3215
            elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path + ".index")):
3216
3217
3218
3219
3220
                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."
                    )
3221
                archive_file = os.path.join(subfolder, pretrained_model_name_or_path + ".index")
3222
                is_local = True
3223
            elif is_remote_url(pretrained_model_name_or_path):
3224
                filename = pretrained_model_name_or_path
3225
                resolved_archive_file = download_url(pretrained_model_name_or_path)
3226
            else:
3227
3228
3229
3230
3231
                # set correct filename
                if from_tf:
                    filename = TF2_WEIGHTS_NAME
                elif from_flax:
                    filename = FLAX_WEIGHTS_NAME
3232
                elif use_safetensors is not False:
3233
                    filename = _add_variant(SAFE_WEIGHTS_NAME, variant)
3234
                else:
3235
                    filename = _add_variant(WEIGHTS_NAME, variant)
3236

3237
3238
                try:
                    # Load from URL or cache if already cached
3239
3240
3241
3242
3243
3244
                    cached_file_kwargs = {
                        "cache_dir": cache_dir,
                        "force_download": force_download,
                        "proxies": proxies,
                        "resume_download": resume_download,
                        "local_files_only": local_files_only,
3245
                        "token": token,
3246
3247
3248
3249
3250
3251
                        "user_agent": user_agent,
                        "revision": revision,
                        "subfolder": subfolder,
                        "_raise_exceptions_for_missing_entries": False,
                        "_commit_hash": commit_hash,
                    }
3252
                    resolved_archive_file = cached_file(pretrained_model_name_or_path, filename, **cached_file_kwargs)
3253

3254
                    # Since we set _raise_exceptions_for_missing_entries=False, we don't get an exception but a None
3255
                    # result when internet is up, the repo and revision exist, but the file does not.
3256
                    if resolved_archive_file is None and filename == _add_variant(SAFE_WEIGHTS_NAME, variant):
3257
3258
                        # Maybe the checkpoint is sharded, we try to grab the index name in this case.
                        resolved_archive_file = cached_file(
3259
3260
3261
                            pretrained_model_name_or_path,
                            _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant),
                            **cached_file_kwargs,
3262
3263
3264
                        )
                        if resolved_archive_file is not None:
                            is_sharded = True
3265
                        elif use_safetensors:
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
                            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`."
                                )
3278
3279
                        else:
                            # This repo has no safetensors file of any kind, we switch to PyTorch.
3280
                            filename = _add_variant(WEIGHTS_NAME, variant)
3281
                            resolved_archive_file = cached_file(
3282
                                pretrained_model_name_or_path, filename, **cached_file_kwargs
3283
                            )
3284
                    if resolved_archive_file is None and filename == _add_variant(WEIGHTS_NAME, variant):
Sylvain Gugger's avatar
Sylvain Gugger committed
3285
                        # Maybe the checkpoint is sharded, we try to grab the index name in this case.
3286
                        resolved_archive_file = cached_file(
3287
3288
3289
                            pretrained_model_name_or_path,
                            _add_variant(WEIGHTS_INDEX_NAME, variant),
                            **cached_file_kwargs,
3290
                        )
3291
3292
3293
                        if resolved_archive_file is not None:
                            is_sharded = True
                    if resolved_archive_file is None:
Sylvain Gugger's avatar
Sylvain Gugger committed
3294
3295
3296
3297
3298
                        # Otherwise, maybe there is a TF or Flax model file.  We try those to give a helpful error
                        # message.
                        has_file_kwargs = {
                            "revision": revision,
                            "proxies": proxies,
3299
                            "token": token,
Sylvain Gugger's avatar
Sylvain Gugger committed
3300
3301
3302
                        }
                        if has_file(pretrained_model_name_or_path, TF2_WEIGHTS_NAME, **has_file_kwargs):
                            raise EnvironmentError(
Sylvain Gugger's avatar
Sylvain Gugger committed
3303
                                f"{pretrained_model_name_or_path} does not appear to have a file named"
3304
3305
                                f" {_add_variant(WEIGHTS_NAME, variant)} but there is a file for TensorFlow weights."
                                " Use `from_tf=True` to load this model from those weights."
Sylvain Gugger's avatar
Sylvain Gugger committed
3306
3307
3308
                            )
                        elif has_file(pretrained_model_name_or_path, FLAX_WEIGHTS_NAME, **has_file_kwargs):
                            raise EnvironmentError(
Sylvain Gugger's avatar
Sylvain Gugger committed
3309
                                f"{pretrained_model_name_or_path} does not appear to have a file named"
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
                                f" {_add_variant(WEIGHTS_NAME, variant)} but there is a file for Flax weights. Use"
                                " `from_flax=True` to load this model from those weights."
                            )
                        elif variant is not None and has_file(
                            pretrained_model_name_or_path, WEIGHTS_NAME, **has_file_kwargs
                        ):
                            raise EnvironmentError(
                                f"{pretrained_model_name_or_path} does not appear to have a file named"
                                f" {_add_variant(WEIGHTS_NAME, variant)} but there is a file without the variant"
                                f" {variant}. Use `variant=None` to load this model from those weights."
Sylvain Gugger's avatar
Sylvain Gugger committed
3320
3321
3322
                            )
                        else:
                            raise EnvironmentError(
3323
3324
3325
                                f"{pretrained_model_name_or_path} does not appear to have a file named"
                                f" {_add_variant(WEIGHTS_NAME, variant)}, {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME} or"
                                f" {FLAX_WEIGHTS_NAME}."
Sylvain Gugger's avatar
Sylvain Gugger committed
3326
                            )
3327
3328
3329
3330
                except EnvironmentError:
                    # Raise any environment error raise by `cached_file`. It will have a helpful error message adapted
                    # to the original exception.
                    raise
3331
                except Exception as e:
3332
                    # For any other exception, we throw a generic error.
3333
                    raise EnvironmentError(
3334
3335
3336
                        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"
3337
3338
                        f" directory containing a file named {_add_variant(WEIGHTS_NAME, variant)},"
                        f" {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME} or {FLAX_WEIGHTS_NAME}."
3339
                    ) from e
3340

3341
            if is_local:
3342
                logger.info(f"loading weights file {archive_file}")
3343
                resolved_archive_file = archive_file
3344
            else:
3345
                logger.info(f"loading weights file {filename} from cache at {resolved_archive_file}")
3346
        else:
thomwolf's avatar
thomwolf committed
3347
            resolved_archive_file = None
3348

Sylvain Gugger's avatar
Sylvain Gugger committed
3349
3350
        # We'll need to download and cache each checkpoint shard if the checkpoint is sharded.
        if is_sharded:
3351
            # rsolved_archive_file becomes a list of files that point to the different checkpoint shards in this case.
Sylvain Gugger's avatar
Sylvain Gugger committed
3352
3353
3354
3355
3356
3357
3358
3359
            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,
3360
                token=token,
Sylvain Gugger's avatar
Sylvain Gugger committed
3361
3362
                user_agent=user_agent,
                revision=revision,
3363
                subfolder=subfolder,
3364
                _commit_hash=commit_hash,
Sylvain Gugger's avatar
Sylvain Gugger committed
3365
3366
            )

3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
        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.")
            else:
                raise ValueError(
                    f"Incompatible safetensors file. File metadata is not ['pt', 'tf', 'flax'] but {metadata.get('format')}"
                )

        from_pt = not (from_tf | from_flax)

3390
3391
        # load pt weights early so that we know which dtype to init the model under
        if from_pt:
3392
            if not is_sharded and state_dict is None:
Sylvain Gugger's avatar
Sylvain Gugger committed
3393
3394
                # Time to load the checkpoint
                state_dict = load_state_dict(resolved_archive_file)
3395

3396
3397
3398
            # 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
3399
            #    weights entry that is of a floating type - we assume all floating dtype weights are of the same dtype
3400
3401
            # we also may have config.torch_dtype available, but we won't rely on it till v5
            dtype_orig = None
3402

3403
3404
3405
            if torch_dtype is not None:
                if isinstance(torch_dtype, str):
                    if torch_dtype == "auto":
3406
3407
3408
                        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
3409
                        else:
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
                            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"
                            )
3422
3423
                    else:
                        raise ValueError(
3424
                            f'`torch_dtype` can be either `torch.dtype` or `"auto"`, but received {torch_dtype}'
3425
3426
3427
                        )
                dtype_orig = cls._set_default_torch_dtype(torch_dtype)

3428
            # Check if `_keep_in_fp32_modules` is not None
3429
3430
            use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and (
                torch_dtype == torch.float16 or load_in_4bit or load_in_8bit
3431
3432
            )

3433
3434
3435
            if is_sharded:
                loaded_state_dict_keys = sharded_metadata["all_checkpoint_keys"]
            else:
3436
                loaded_state_dict_keys = list(state_dict.keys())
3437
3438
3439
3440
            if low_cpu_mem_usage or (use_keep_in_fp32_modules and is_accelerate_available()):
                # In case some weights need to be kept in float32 and accelerate is not installed,
                # we later on want to take the path where state_dict is not None, that is the one
                # that do not require accelerate.
3441
                state_dict = None
3442

3443
3444
        config.name_or_path = pretrained_model_name_or_path

3445
        # Instantiate model.
3446
3447
        init_contexts = [no_init_weights(_enable=_fast_init)]

3448
3449
3450
3451
        if is_deepspeed_zero3_enabled():
            import deepspeed

            logger.info("Detected DeepSpeed ZeRO-3: activating zero.init() for this model")
3452
            init_contexts = [deepspeed.zero.Init(config_dict_or_path=deepspeed_config())] + init_contexts
3453
        elif load_in_8bit or load_in_4bit or low_cpu_mem_usage:
3454
3455
            init_contexts.append(init_empty_weights())

3456
3457
3458
3459
        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
        )
3460

3461
        with ContextManagers(init_contexts):
3462
            # Let's make sure we don't run the init function of buffer modules
3463
3464
            model = cls(config, *model_args, **model_kwargs)

3465
3466
3467
        # make sure we use the model's config since the __init__ call might have copied it
        config = model.config

3468
3469
        # Check first if we are `from_pt`
        if use_keep_in_fp32_modules:
3470
            if is_accelerate_available() and not is_deepspeed_zero3_enabled():
3471
                low_cpu_mem_usage = True
3472
3473
3474
3475
            keep_in_fp32_modules = model._keep_in_fp32_modules
        else:
            keep_in_fp32_modules = []

3476
        if load_in_8bit or load_in_4bit:
3477
            from .integrations import get_keys_to_not_convert, replace_with_bnb_linear
3478

3479
            llm_int8_skip_modules = quantization_config.llm_int8_skip_modules
3480
            load_in_8bit_fp32_cpu_offload = quantization_config.llm_int8_enable_fp32_cpu_offload
3481
3482
3483
3484
            if load_in_8bit:
                logger.info("Detected 8-bit loading: activating 8-bit loading for this model")
            else:
                logger.info("Detected 4-bit loading: activating 4-bit loading for this model")
3485

3486
            # We keep some modules such as the lm_head in their original dtype for numerical stability reasons
3487
            if llm_int8_skip_modules is None:
3488
3489
                modules_to_not_convert = get_keys_to_not_convert(model)
            else:
3490
                modules_to_not_convert = llm_int8_skip_modules
3491
3492
3493
3494
3495
3496

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

            modules_to_not_convert.extend(keep_in_fp32_modules)

3497
3498
3499
3500
3501
3502
3503
            # Extend the modules to not convert to keys that are supposed to be offloaded to `cpu` or `disk`
            if isinstance(device_map, dict) and len(device_map.keys()) > 1:
                keys_on_cpu = [key for key, value in device_map.items() if value in ["disk", "cpu"]]

                if len(keys_on_cpu) > 0 and not load_in_8bit_fp32_cpu_offload:
                    raise ValueError(
                        "If you want to offload some keys to `cpu` or `disk`, you need to set "
Younes Belkada's avatar
Younes Belkada committed
3504
                        "`llm_int8_enable_fp32_cpu_offload=True`. Note that these modules will not be "
3505
3506
3507
3508
3509
                        " converted to 8-bit but kept in 32-bit."
                    )

                modules_to_not_convert.extend(keys_on_cpu)

3510
            supports_4bit = version.parse(importlib.metadata.version("bitsandbytes")) >= version.parse("0.39.0")
3511
3512
3513
3514
3515
3516
3517
3518
3519

            if load_in_4bit and not supports_4bit:
                raise ValueError(
                    "You have a version of `bitsandbytes` that is not compatible with 4bit inference and training"
                    " make sure you have the latest version of `bitsandbytes` installed"
                )

            model = replace_with_bnb_linear(
                model, modules_to_not_convert=modules_to_not_convert, quantization_config=quantization_config
3520
            )
3521
            # training in 8-bit is only available in 0.37.0+
3522
            model._is_quantized_training_enabled = version.parse(
3523
                importlib.metadata.version("bitsandbytes")
3524
            ) >= version.parse("0.37.0")
3525

3526
            config.quantization_config = quantization_config
3527
3528
            model.is_8bit_serializable = is_8bit_serializable

3529
3530
        if load_in_8bit and torch_dtype is None:
            logger.warning(
3531
                "You are loading your model in 8bit but you did not specify a `torch_dtype` attribute. "
3532
3533
                "All non-linear modules will be loaded in full precision."
                " If you want to load the other modules in other precision, please specify a `torch_dtype` attribute."
3534
            )
Marc Sun's avatar
Marc Sun committed
3535
3536
3537
        if quantization_method_from_config == QuantizationMethod.GPTQ:
            model = quantizer.convert_model(model)
            model._is_quantized_training_enabled = True
3538
        elif quantization_method_from_config == QuantizationMethod.AWQ:
3539
            from .integrations import fuse_awq_modules, get_keys_to_not_convert, replace_with_awq_linear
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556

            modules_to_not_convert = get_keys_to_not_convert(model)

            if quantization_config is None:
                quantization_config = AwqConfig.from_dict(config.quantization_config)

            model, has_been_replaced = replace_with_awq_linear(
                model, quantization_config=quantization_config, modules_to_not_convert=modules_to_not_convert
            )
            model._is_quantized_training_enabled = False

            if not has_been_replaced:
                logger.warning(
                    "You are loading an AWQ model but no linear modules were found in your model."
                    " Please double check your model architecture, or submit an issue on github if you think this is"
                    " a bug."
                )
Marc Sun's avatar
Marc Sun committed
3557
3558
3559
3560
3561
3562
3563

        if quantization_method_from_config is not None:
            model.quantization_method = quantization_method_from_config
        elif quantization_method_from_args is not None:
            model.quantization_method = quantization_method_from_args
        if hasattr(model, "quantization_method"):
            model.is_quantized = True
3564

3565
3566
3567
3568
3569
3570
            # 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

3571
        if isinstance(device_map, str):
3572
            special_dtypes = {}
3573
            if load_in_8bit or load_in_4bit:
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
                special_dtypes.update(
                    {
                        name: torch_dtype
                        for name, _ in model.named_parameters()
                        if any(m in name for m in modules_to_not_convert)
                    }
                )

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

3590
3591
3592
            target_dtype = torch_dtype

            if load_in_4bit:
3593
                if version.parse(importlib.metadata.version("accelerate")) > version.parse("0.19.0"):
3594
3595
3596
3597
3598
3599
                    from accelerate.utils import CustomDtype

                    target_dtype = CustomDtype.INT4
                else:
                    raise ValueError(
                        "You are using `device_map='auto'` on a 4bit loaded version of the model. To automatically compute"
3600
3601
                        " the appropriate device map, you should upgrade your `accelerate` library, "
                        "`pip install --upgrade accelerate` or install it from source to support fp4 auto device map "
3602
3603
3604
3605
3606
                        "calculation. You may encounter unexpected behavior, or pass your own device map"
                    )
            elif load_in_8bit:
                target_dtype = torch.int8

Marc Sun's avatar
Marc Sun committed
3607
            no_split_modules = model._get_no_split_modules(device_map)
3608
3609
3610
3611
3612
            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'."
                )
3613

3614
            device_map_kwargs = {"no_split_module_classes": no_split_modules}
3615
            if "special_dtypes" in inspect.signature(infer_auto_device_map).parameters:
3616
                device_map_kwargs["special_dtypes"] = special_dtypes
3617
            elif len(special_dtypes) > 0:
3618
                logger.warning(
3619
3620
3621
                    "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`)."
                )
3622
            if device_map != "sequential":
3623
3624
                max_memory = get_balanced_memory(
                    model,
3625
                    dtype=target_dtype,
3626
                    low_zero=(device_map == "balanced_low_0"),
3627
                    max_memory=max_memory,
3628
                    **device_map_kwargs,
3629
                )
Marc Sun's avatar
Marc Sun committed
3630
3631
3632
3633
3634
            else:
                max_memory = get_max_memory(max_memory)
            if getattr(model, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
                # need more space for buffers that are created during quantization
                max_memory = {key: val * 0.90 for key, val in max_memory.items()}
3635
            device_map_kwargs["max_memory"] = max_memory
Marc Sun's avatar
Marc Sun committed
3636

3637
3638
            # Make sure tied weights are tied before creating the device map.
            model.tie_weights()
3639
            device_map = infer_auto_device_map(model, dtype=target_dtype, **device_map_kwargs)
3640

3641
            if load_in_8bit or load_in_4bit:
3642
                # The LM head / tied weights or any last module can stay on disk / CPU
3643
                device_map_without_lm_head = {
3644
                    key: device_map[key] for key in device_map.keys() if key not in modules_to_not_convert
3645
3646
                }
                if "cpu" in device_map_without_lm_head.values() or "disk" in device_map_without_lm_head.values():
3647
3648
3649
                    raise ValueError(
                        """
                        Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit
3650
3651
3652
3653
3654
                        the quantized model. If you want to dispatch the model on the CPU or the disk while keeping
                        these modules in 32-bit, you need to set `load_in_8bit_fp32_cpu_offload=True` and pass a custom
                        `device_map` to `from_pretrained`. Check
                        https://huggingface.co/docs/transformers/main/en/main_classes/quantization#offload-between-cpu-and-gpu
                        for more details.
3655
3656
                        """
                    )
3657
3658
                del device_map_without_lm_head

3659
3660
3661
3662
        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
3663
            check_tied_parameters_on_same_device(tied_params, device_map)
3664

3665
        if from_tf:
3666
            if resolved_archive_file.endswith(".index"):
3667
3668
3669
3670
3671
                # 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:
3672
                    from .modeling_tf_pytorch_utils import load_tf2_checkpoint_in_pytorch_model
3673

Yih-Dar's avatar
Yih-Dar committed
3674
3675
3676
                    model, loading_info = load_tf2_checkpoint_in_pytorch_model(
                        model, resolved_archive_file, allow_missing_keys=True, output_loading_info=True
                    )
3677
                except ImportError:
3678
                    logger.error(
Sylvain Gugger's avatar
Sylvain Gugger committed
3679
3680
3681
                        "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."
3682
                    )
3683
                    raise
3684
3685
3686
3687
3688
3689
3690
        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
3691
3692
3693
                    "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."
3694
3695
                )
                raise
3696
        elif from_pt:
3697
3698
3699
            # restore default dtype
            if dtype_orig is not None:
                torch.set_default_dtype(dtype_orig)
Sylvain Gugger's avatar
Sylvain Gugger committed
3700
3701
3702
3703
3704
3705
3706
3707
            (
                model,
                missing_keys,
                unexpected_keys,
                mismatched_keys,
                offload_index,
                error_msgs,
            ) = cls._load_pretrained_model(
3708
3709
3710
3711
3712
3713
3714
3715
3716
                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,
3717
3718
3719
3720
                device_map=device_map,
                offload_folder=offload_folder,
                offload_state_dict=offload_state_dict,
                dtype=torch_dtype,
Marc Sun's avatar
Marc Sun committed
3721
                is_quantized=(getattr(model, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES),
3722
                keep_in_fp32_modules=keep_in_fp32_modules,
3723
            )
3724

3725
        model.is_loaded_in_4bit = load_in_4bit
Younes Belkada's avatar
Younes Belkada committed
3726
        model.is_loaded_in_8bit = load_in_8bit
3727

3728
3729
        # make sure token embedding weights are still tied if needed
        model.tie_weights()
3730

3731
        # Set model in evaluation mode to deactivate DropOut modules by default
3732
3733
        model.eval()

3734
        # If it is a model with generation capabilities, attempt to load the generation config
3735
        if model.can_generate() and pretrained_model_name_or_path is not None:
3736
3737
3738
3739
3740
3741
3742
3743
            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,
3744
                    token=token,
3745
3746
3747
3748
3749
3750
                    revision=revision,
                    subfolder=subfolder,
                    _from_auto=from_auto_class,
                    _from_pipeline=from_pipeline,
                    **kwargs,
                )
3751
            except OSError:
3752
3753
3754
3755
3756
                logger.info(
                    "Generation config file not found, using a generation config created from the model config."
                )
                pass

3757
3758
3759
3760
3761
3762
3763
3764
        if (
            quantization_config is not None
            and quantization_config.quant_method == QuantizationMethod.AWQ
            and quantization_config.do_fuse
        ):
            model = fuse_awq_modules(model, config.quantization_config)
            model._awq_is_fused = True

3765
3766
        # Dispatch model with hooks on all devices if necessary
        if device_map is not None:
3767
3768
3769
3770
3771
            device_map_kwargs = {
                "device_map": device_map,
                "offload_dir": offload_folder,
                "offload_index": offload_index,
            }
3772
            if "skip_keys" in inspect.signature(dispatch_model).parameters:
3773
3774
                device_map_kwargs["skip_keys"] = model._skip_keys_device_placement
            dispatch_model(model, **device_map_kwargs)
3775

Marc Sun's avatar
Marc Sun committed
3776
3777
3778
3779
3780
3781
        if quantization_method_from_args == QuantizationMethod.GPTQ:
            if quantization_config.tokenizer is None:
                quantization_config.tokenizer = pretrained_model_name_or_path
            if cls.main_input_name != "input_ids":
                raise RuntimeError("We can only quantize pure text model.")
            quantizer.quantize_model(model, quantization_config.tokenizer)
Marc Sun's avatar
Marc Sun committed
3782
            config.quantization_config = GPTQConfig.from_dict_optimum(quantizer.to_dict())
Marc Sun's avatar
Marc Sun committed
3783
3784
3785
3786
            model._is_quantized_training_enabled = True
        if quantization_method_from_config == QuantizationMethod.GPTQ:
            model = quantizer.post_init_model(model)

3787
        if _adapter_model_path is not None:
3788
            model.load_adapter(
3789
                _adapter_model_path,
3790
3791
                adapter_name=adapter_name,
                token=token,
3792
                adapter_kwargs=adapter_kwargs,
3793
3794
            )

thomwolf's avatar
thomwolf committed
3795
        if output_loading_info:
Yih-Dar's avatar
Yih-Dar committed
3796
3797
3798
3799
3800
3801
3802
            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
3803
3804
            return model, loading_info

3805
3806
        return model

3807
    @classmethod
Sylvain Gugger's avatar
Sylvain Gugger committed
3808
3809
3810
3811
    def _load_pretrained_model(
        cls,
        model,
        state_dict,
3812
        loaded_keys,
Sylvain Gugger's avatar
Sylvain Gugger committed
3813
3814
3815
3816
3817
        resolved_archive_file,
        pretrained_model_name_or_path,
        ignore_mismatched_sizes=False,
        sharded_metadata=None,
        _fast_init=True,
3818
        low_cpu_mem_usage=False,
3819
3820
        device_map=None,
        offload_folder=None,
3821
        offload_state_dict=None,
3822
        dtype=None,
3823
        is_quantized=False,
3824
        keep_in_fp32_modules=None,
3825
    ):
Sylvain Gugger's avatar
Sylvain Gugger committed
3826
        is_safetensors = False
3827
        if is_quantized:
3828
            from .integrations import set_module_quantized_tensor_to_device
3829

Sylvain Gugger's avatar
Sylvain Gugger committed
3830
        if device_map is not None and "disk" in device_map.values():
Sylvain Gugger's avatar
Sylvain Gugger committed
3831
3832
3833
3834
3835
            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
3836
3837
                raise ValueError(
                    "The current `device_map` had weights offloaded to the disk. Please provide an `offload_folder`"
Sylvain Gugger's avatar
Sylvain Gugger committed
3838
3839
                    " 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
3840
                )
Sylvain Gugger's avatar
Sylvain Gugger committed
3841
3842
            if offload_folder is not None:
                os.makedirs(offload_folder, exist_ok=True)
3843
3844
3845
            if offload_state_dict is None:
                offload_state_dict = True

3846
        is_sharded_safetensors = is_safetensors and sharded_metadata is not None
Patrick von Platen's avatar
Patrick von Platen committed
3847
3848
3849
3850

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

3851
        # Retrieve missing & unexpected_keys
3852
3853
        model_state_dict = model.state_dict()
        expected_keys = list(model_state_dict.keys())
3854
3855
        prefix = model.base_model_prefix

Sylvain Gugger's avatar
Sylvain Gugger committed
3856
3857
3858
3859
3860
3861
3862
        def _fix_key(key):
            if "beta" in key:
                return key.replace("beta", "bias")
            if "gamma" in key:
                return key.replace("gamma", "weight")
            return key

3863
        original_loaded_keys = loaded_keys
Sylvain Gugger's avatar
Sylvain Gugger committed
3864
3865
        loaded_keys = [_fix_key(key) for key in loaded_keys]

3866
3867
3868
3869
3870
3871
        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
3872
3873
3874

        # key re-naming operations are never done on the keys
        # that are loaded, but always on the keys of the newly initialized model
3875
3876
        remove_prefix_from_model = not has_prefix_module and expects_prefix_module
        add_prefix_to_model = has_prefix_module and not expects_prefix_module
3877

3878
        if remove_prefix_from_model:
3879
3880
3881
            _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]
3882
        elif add_prefix_to_model:
3883
3884
3885
            expected_keys = [".".join([prefix, s]) for s in expected_keys]

        missing_keys = list(set(expected_keys) - set(loaded_keys))
Sylvain Gugger's avatar
Sylvain Gugger committed
3886
3887
3888
3889
3890
3891
3892
3893
3894
        unexpected_keys = set(loaded_keys) - set(expected_keys)
        # Remove nonpersistent buffers from unexpected keys: they are not in the state dict but will be in the model
        # buffers
        model_buffers = {n for n, _ in model.named_buffers()}
        if remove_prefix_from_model:
            model_buffers = {key[len(_prefix) :] if key.startswith(_prefix) else key for key in model_buffers}
        elif add_prefix_to_model:
            model_buffers = {".".join([prefix, key]) for key in model_buffers}
        unexpected_keys = list(unexpected_keys - model_buffers)
3895

3896
3897
        model.tie_weights()
        if device_map is None and not is_fsdp_enabled():
3898
3899
3900
3901
            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
3902

3903
3904
3905
3906
3907
            # 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
3908
3909

        for group in tied_params:
Sylvain Gugger's avatar
Sylvain Gugger committed
3910
3911
3912
3913
            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
3914
3915
3916
            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]
3917

3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
        # 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]

3928
3929
3930
3931
        # 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
3932
3933
                if key in list(model_state_dict.keys()):
                    key = key
3934
3935
                elif f"{prefix}.{key}" in list(model_state_dict.keys()):
                    key = f"{prefix}.{key}"
Susnato Dhar's avatar
Susnato Dhar committed
3936
                elif key.startswith(prefix) and ".".join(key.split(".")[1:]) in list(model_state_dict.keys()):
3937
3938
                    key = ".".join(key.split(".")[1:])
                param = model_state_dict[key]
3939
3940
3941
3942
3943
3944

                # upcast in fp32 if any
                target_dtype = dtype
                if (
                    keep_in_fp32_modules is not None
                    and dtype == torch.float16
3945
3946
3947
                    and any(
                        module_to_keep_in_fp32 in key.split(".") for module_to_keep_in_fp32 in keep_in_fp32_modules
                    )
3948
3949
3950
                ):
                    target_dtype = torch.float32

3951
                if param.device == torch.device("meta"):
3952
                    if not (is_quantized):
3953
                        set_module_tensor_to_device(model, key, "cpu", torch.empty(*param.size(), dtype=target_dtype))
3954
                    else:
3955
                        set_module_quantized_tensor_to_device(
3956
3957
                            model, key, "cpu", torch.empty(*param.size(), dtype=target_dtype)
                        )
3958
3959

        # retrieve unintialized modules and initialize before maybe overriding that with the pretrained weights.
3960
        if _fast_init:
3961
3962
3963
3964
3965
3966
3967
3968
            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
                set_initialized_submodules(model, _loaded_keys)
3969
3970
            # This will only initialize submodules that are not marked as initialized by the line above.
            model.apply(model._initialize_weights)
3971

3972
3973
3974
        # Set some modules to fp32 if any
        if keep_in_fp32_modules is not None:
            for name, param in model.named_parameters():
3975
                if any(module_to_keep_in_fp32 in name.split(".") for module_to_keep_in_fp32 in keep_in_fp32_modules):
3976
3977
                    # param = param.to(torch.float32) does not work here as only in the local scope.
                    param.data = param.data.to(torch.float32)
3978

3979
3980
3981
        # Make sure we are able to load base models as well as derived models (with heads)
        start_prefix = ""
        model_to_load = model
3982
        if len(cls.base_model_prefix) > 0 and not hasattr(model, cls.base_model_prefix) and has_prefix_module:
3983
            start_prefix = cls.base_model_prefix + "."
3984
        if len(cls.base_model_prefix) > 0 and hasattr(model, cls.base_model_prefix) and not has_prefix_module:
3985
            model_to_load = getattr(model, cls.base_model_prefix)
Sylvain Gugger's avatar
Sylvain Gugger committed
3986
3987
            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):
3988
                raise ValueError(
3989
                    "The state dictionary of the model you are trying to load is corrupted. Are you sure it was "
3990
3991
                    "properly saved?"
                )
3992
3993
            if device_map is not None:
                device_map = {k.replace(f"{cls.base_model_prefix}.", ""): v for k, v in device_map.items()}
3994

3995
3996
3997
3998
3999
4000
4001
4002
        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
4003
4004
4005
            mismatched_keys = []
            if ignore_mismatched_sizes:
                for checkpoint_key in loaded_keys:
4006
4007
4008
                    # 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
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
                    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
                    ):
                        mismatched_keys.append(
                            (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
                        )
                        del state_dict[checkpoint_key]
4025
4026
            return mismatched_keys

4027
4028
4029
4030
        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
4031
        if device_map is not None and is_safetensors:
4032
            param_device_map = expand_device_map(device_map, original_loaded_keys, start_prefix)
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
            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
4043
            offload_index = {
4044
                p[len(start_prefix) :]: {"safetensors_file": f, "weight_name": p, "dtype": str_dtype}
4045
                for p, f in weight_map.items()
4046
                if p.startswith(start_prefix) and param_device_map[p[len(start_prefix) :]] == "disk"
Sylvain Gugger's avatar
Sylvain Gugger committed
4047
4048
            }

4049
4050
4051
4052
4053
        if state_dict is not None:
            # Whole checkpoint
            mismatched_keys = _find_mismatched_keys(
                state_dict,
                model_state_dict,
4054
                original_loaded_keys,
4055
4056
4057
4058
                add_prefix_to_model,
                remove_prefix_from_model,
                ignore_mismatched_sizes,
            )
Sylvain Gugger's avatar
Sylvain Gugger committed
4059
            error_msgs = _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
Sylvain Gugger's avatar
Sylvain Gugger committed
4060
            offload_index = None
Sylvain Gugger's avatar
Sylvain Gugger committed
4061
        else:
4062
4063
            # Sharded checkpoint or whole but low_cpu_mem_usage==True

Sylvain Gugger's avatar
Sylvain Gugger committed
4064
4065
4066
4067
4068
            # 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 = []
4069
            mismatched_keys = []
Sylvain Gugger's avatar
Sylvain Gugger committed
4070
4071
            if not is_safetensors:
                offload_index = {} if device_map is not None and "disk" in device_map.values() else None
4072
4073
4074
4075
4076
4077
4078
            if offload_state_dict:
                state_dict_folder = tempfile.mkdtemp()
                state_dict_index = {}
            else:
                state_dict_folder = None
                state_dict_index = None

4079
            if is_sharded_safetensors:
4080
4081
4082
                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
4083
4084
4085
4086
                disk_only_shard_files = [os.path.join(folder, f) for f in disk_only_shard_files]
            else:
                disk_only_shard_files = []

4087
4088
            if len(resolved_archive_file) > 1:
                resolved_archive_file = logging.tqdm(resolved_archive_file, desc="Loading checkpoint shards")
Sylvain Gugger's avatar
Sylvain Gugger committed
4089
            for shard_file in resolved_archive_file:
Sylvain Gugger's avatar
Sylvain Gugger committed
4090
4091
4092
                # 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
Sylvain Gugger's avatar
Sylvain Gugger committed
4093
                state_dict = load_state_dict(shard_file)
4094

Sylvain Gugger's avatar
Sylvain Gugger committed
4095
4096
                # Mistmatched keys contains tuples key/shape1/shape2 of weights in the checkpoint that have a shape not
                # matching the weights in the model.
4097
4098
4099
                mismatched_keys += _find_mismatched_keys(
                    state_dict,
                    model_state_dict,
4100
                    original_loaded_keys,
4101
4102
4103
4104
                    add_prefix_to_model,
                    remove_prefix_from_model,
                    ignore_mismatched_sizes,
                )
4105
                if low_cpu_mem_usage:
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
                    if is_fsdp_enabled() and not is_local_dist_rank_0():
                        for key, param in model_to_load.state_dict().items():
                            if param.device == torch.device("meta"):
                                if not (is_quantized):
                                    set_module_tensor_to_device(
                                        model_to_load, key, "cpu", torch.empty(*param.size(), dtype=dtype)
                                    )
                                else:
                                    set_module_quantized_tensor_to_device(
                                        model_to_load, key, "cpu", torch.empty(*param.size(), dtype=dtype)
                                    )
                    else:
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
                        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,
                            is_quantized=is_quantized,
                            is_safetensors=is_safetensors,
                            keep_in_fp32_modules=keep_in_fp32_modules,
                        )
                        error_msgs += new_error_msgs
4135
4136
                else:
                    error_msgs += _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
4137

4138
4139
4140
4141
                # force memory release
                del state_dict
                gc.collect()

4142
            if offload_index is not None and len(offload_index) > 0:
Sylvain Gugger's avatar
Sylvain Gugger committed
4143
4144
4145
                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
4146
4147
4148
4149
4150
4151
                    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
4152
                    offload_index = {f"{prefix}.{key}": value for key, value in offload_index.items()}
Sylvain Gugger's avatar
Sylvain Gugger committed
4153
4154
4155
                if not is_safetensors:
                    save_offload_index(offload_index, offload_folder)
                    offload_index = None
4156
4157
4158

            if offload_state_dict:
                # Load back temporarily offloaded state dict
4159
                load_offloaded_weights(model_to_load, state_dict_index, state_dict_folder)
4160
4161
                shutil.rmtree(state_dict_folder)

4162
4163
        if len(error_msgs) > 0:
            error_msg = "\n\t".join(error_msgs)
4164
4165
4166
4167
            if "size mismatch" in error_msg:
                error_msg += (
                    "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
                )
4168
4169
            raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")

4170
        if is_quantized:
4171
4172
4173
            unexpected_keys = [elem for elem in unexpected_keys if "SCB" not in elem]
            missing_keys = [elem for elem in missing_keys if "SCB" not in elem]

4174
        if len(unexpected_keys) > 0:
Sylvain Gugger's avatar
Sylvain Gugger committed
4175
            archs = [] if model.config.architectures is None else model.config.architectures
4176
            warner = logger.warning if model.__class__.__name__ in archs else logger.info
Sylvain Gugger's avatar
Sylvain Gugger committed
4177
            warner(
Sylvain Gugger's avatar
Sylvain Gugger committed
4178
4179
4180
4181
4182
4183
4184
                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)."
4185
4186
4187
4188
4189
            )
        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
4190
4191
4192
                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."
4193
            )
4194
        elif len(mismatched_keys) == 0:
4195
            logger.info(
Sylvain Gugger's avatar
Sylvain Gugger committed
4196
4197
4198
4199
                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."
4200
            )
4201
4202
4203
4204
4205
4206
4207
4208
        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
4209
4210
4211
4212
                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."
4213
            )
4214

Sylvain Gugger's avatar
Sylvain Gugger committed
4215
        return model, missing_keys, unexpected_keys, mismatched_keys, offload_index, error_msgs
4216
4217

    def retrieve_modules_from_names(self, names, add_prefix=False, remove_prefix=False):
4218
        module_keys = {".".join(key.split(".")[:-1]) for key in names}
4219

Patrick von Platen's avatar
Patrick von Platen committed
4220
4221
        # torch.nn.ParameterList is a special case where two parameter keywords
        # are appended to the module name, *e.g.* bert.special_embeddings.0
4222
        module_keys = module_keys.union(
4223
            {".".join(key.split(".")[:-2]) for key in names if len(key) > 0 and key[-1].isdigit()}
4224
        )
Patrick von Platen's avatar
Patrick von Platen committed
4225

4226
4227
4228
4229
        retrieved_modules = []
        # retrieve all modules that has at least one missing weight name
        for name, module in self.named_modules():
            if remove_prefix:
4230
4231
                _prefix = f"{self.base_model_prefix}."
                name = name[len(_prefix) :] if name.startswith(_prefix) else name
4232
            elif add_prefix:
Patrick von Platen's avatar
Patrick von Platen committed
4233
                name = ".".join([self.base_model_prefix, name]) if len(name) > 0 else self.base_model_prefix
4234
4235
4236
4237
4238
4239

            if name in module_keys:
                retrieved_modules.append(module)

        return retrieved_modules

4240
    @staticmethod
4241
    def _load_pretrained_model_low_mem(model, loaded_state_dict_keys, resolved_archive_file, start_prefix=""):
4242
4243
4244
        """
        This is an experimental function that loads the model using ~1.x model size CPU memory

4245
        Before you call it do:
4246

4247
        1. save which state_dict keys are available
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
        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

        Currently, it doesn't handle missing_keys, unexpected_keys, mismatched_keys. It can't handle deepspeed.
        """

4259
4260
4261
4262
        _move_model_to_meta(model, loaded_state_dict_keys, start_prefix)
        state_dict = load_state_dict(resolved_archive_file)
        error_msgs = _load_state_dict_into_meta_model(model, state_dict, loaded_state_dict_keys, start_prefix)
        return error_msgs
4263

4264
4265
4266
4267
4268
4269
    @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.

4270
4271
4272
4273
4274
4275
        <Tip warning={true}>

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

        </Tip>

4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
        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

4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
    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)

4340
4341
4342
4343
    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.
        """
4344
4345

        # Skip the check during tracing.
4346
        if is_torch_fx_proxy(input_ids) or torch.jit.is_tracing() or is_torchdynamo_compiling():
4347
4348
            return

4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
        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)

thomwolf's avatar
thomwolf committed
4375

4376
PreTrainedModel.push_to_hub = copy_func(PreTrainedModel.push_to_hub)
4377
4378
4379
4380
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"
    )
4381
4382


thomwolf's avatar
thomwolf committed
4383
class PoolerStartLogits(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4384
4385
    """
    Compute SQuAD start logits from sequence hidden states.
4386

Sylvain Gugger's avatar
Sylvain Gugger committed
4387
    Args:
4388
4389
        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
4390
4391
4392
    """

    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
4393
        super().__init__()
thomwolf's avatar
thomwolf committed
4394
4395
        self.dense = nn.Linear(config.hidden_size, 1)

Sylvain Gugger's avatar
Sylvain Gugger committed
4396
4397
4398
4399
4400
    def forward(
        self, hidden_states: torch.FloatTensor, p_mask: Optional[torch.FloatTensor] = None
    ) -> torch.FloatTensor:
        """
        Args:
4401
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4402
                The final hidden states of the model.
4403
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4404
4405
                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
4406
4407

        Returns:
4408
            `torch.FloatTensor`: The start logits for SQuAD.
thomwolf's avatar
thomwolf committed
4409
        """
thomwolf's avatar
thomwolf committed
4410
4411
4412
        x = self.dense(hidden_states).squeeze(-1)

        if p_mask is not None:
Lysandre Debut's avatar
Lysandre Debut committed
4413
            if get_parameter_dtype(self) == torch.float16:
4414
4415
4416
                x = x * (1 - p_mask) - 65500 * p_mask
            else:
                x = x * (1 - p_mask) - 1e30 * p_mask
thomwolf's avatar
thomwolf committed
4417
4418
4419
4420
4421
4422

        return x


class PoolerEndLogits(nn.Module):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
4423
    Compute SQuAD end logits from sequence hidden states.
4424

Sylvain Gugger's avatar
Sylvain Gugger committed
4425
    Args:
4426
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
4427
4428
            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
4429
4430
4431
    """

    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
4432
        super().__init__()
thomwolf's avatar
thomwolf committed
4433
4434
4435
4436
4437
        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
4438
4439
4440
4441
4442
4443
4444
4445
4446
    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:
4447
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4448
                The final hidden states of the model.
4449
            start_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4450
                The hidden states of the first tokens for the labeled span.
4451
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4452
                The position of the first token for the labeled span.
4453
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4454
4455
                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
4456

4457
        <Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
4458

Stas Bekman's avatar
Stas Bekman committed
4459
4460
        One of `start_states` or `start_positions` should be not `None`. If both are set, `start_positions` overrides
        `start_states`.
4461
4462

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

        Returns:
4465
            `torch.FloatTensor`: The end logits for SQuAD.
thomwolf's avatar
thomwolf committed
4466
        """
4467
4468
4469
        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
4470
        if start_positions is not None:
4471
            slen, hsz = hidden_states.shape[-2:]
4472
4473
4474
            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
4475
4476
4477
4478
4479
4480
4481

        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
4482
            if get_parameter_dtype(self) == torch.float16:
4483
4484
4485
                x = x * (1 - p_mask) - 65500 * p_mask
            else:
                x = x * (1 - p_mask) - 1e30 * p_mask
thomwolf's avatar
thomwolf committed
4486
4487
4488
4489
4490

        return x


class PoolerAnswerClass(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4491
4492
4493
4494
    """
    Compute SQuAD 2.0 answer class from classification and start tokens hidden states.

    Args:
4495
4496
        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
4497
    """
4498

thomwolf's avatar
thomwolf committed
4499
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
4500
        super().__init__()
thomwolf's avatar
thomwolf committed
4501
4502
4503
4504
        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
4505
4506
4507
4508
4509
4510
4511
    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:
4512
4513
        """
        Args:
4514
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4515
                The final hidden states of the model.
4516
            start_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4517
                The hidden states of the first tokens for the labeled span.
4518
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4519
                The position of the first token for the labeled span.
4520
4521
4522
4523
            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
4524

Stas Bekman's avatar
Stas Bekman committed
4525
4526
        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
4527

4528
        </Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
4529
4530

        Returns:
4531
            `torch.FloatTensor`: The SQuAD 2.0 answer class.
thomwolf's avatar
thomwolf committed
4532
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
4533
        # No dependency on end_feature so that we can obtain one single `cls_logits` for each sample.
4534
        hsz = hidden_states.shape[-1]
4535
4536
4537
        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
4538
        if start_positions is not None:
4539
4540
            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
4541
4542

        if cls_index is not None:
4543
4544
            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
4545
        else:
4546
            cls_token_state = hidden_states[:, -1, :]  # shape (bsz, hsz)
thomwolf's avatar
thomwolf committed
4547
4548
4549
4550
4551
4552
4553
4554

        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


4555
4556
4557
@dataclass
class SquadHeadOutput(ModelOutput):
    """
4558
    Base class for outputs of question answering models using a [`~modeling_utils.SQuADHead`].
4559
4560

    Args:
4561
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned if both `start_positions` and `end_positions` are provided):
Sylvain Gugger's avatar
Sylvain Gugger committed
4562
4563
            Classification loss as the sum of start token, end token (and is_impossible if provided) classification
            losses.
4564
        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):
4565
            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
4566
        start_top_index (`torch.LongTensor` of shape `(batch_size, config.start_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
4567
            Indices for the top config.start_n_top start token possibilities (beam-search).
4568
4569
        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
4570
            (beam-search).
4571
4572
4573
4574
        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.
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585

    """

    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
4586
class SQuADHead(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4587
4588
    r"""
    A SQuAD head inspired by XLNet.
4589

Sylvain Gugger's avatar
Sylvain Gugger committed
4590
    Args:
4591
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
4592
4593
            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
4594
    """
4595

thomwolf's avatar
thomwolf committed
4596
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
4597
        super().__init__()
thomwolf's avatar
thomwolf committed
4598
4599
4600
4601
4602
4603
4604
        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
4605
    @replace_return_docstrings(output_type=SquadHeadOutput, config_class=PretrainedConfig)
4606
    def forward(
4607
        self,
Sylvain Gugger's avatar
Sylvain Gugger committed
4608
4609
4610
4611
4612
4613
        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,
4614
        return_dict: bool = False,
Sylvain Gugger's avatar
Sylvain Gugger committed
4615
4616
    ) -> Union[SquadHeadOutput, Tuple[torch.FloatTensor]]:
        """
Lysandre's avatar
Lysandre committed
4617
        Args:
4618
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Lysandre's avatar
Lysandre committed
4619
                Final hidden states of the model on the sequence tokens.
4620
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Lysandre's avatar
Lysandre committed
4621
                Positions of the first token for the labeled span.
4622
            end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Lysandre's avatar
Lysandre committed
4623
                Positions of the last token for the labeled span.
4624
4625
4626
            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
4627
                Whether the question has a possible answer in the paragraph or not.
4628
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
4629
4630
                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
                should be masked.
4631
            return_dict (`bool`, *optional*, defaults to `False`):
4632
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
Sylvain Gugger's avatar
Sylvain Gugger committed
4633

Lysandre's avatar
Lysandre committed
4634
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
4635
        """
thomwolf's avatar
thomwolf committed
4636
        start_logits = self.start_logits(hidden_states, p_mask=p_mask)
thomwolf's avatar
thomwolf committed
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659

        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
4660

4661
            return SquadHeadOutput(loss=total_loss) if return_dict else (total_loss,)
thomwolf's avatar
thomwolf committed
4662
4663
4664
4665

        else:
            # during inference, compute the end logits based on beam search
            bsz, slen, hsz = hidden_states.size()
4666
            start_log_probs = nn.functional.softmax(start_logits, dim=-1)  # shape (bsz, slen)
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677

            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
4678
4679
            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)
4680
            end_log_probs = nn.functional.softmax(end_logits, dim=1)  # shape (bsz, slen, start_n_top)
thomwolf's avatar
thomwolf committed
4681

4682
4683
4684
            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
4685
4686
4687
4688
4689
4690
            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)

4691
            if not return_dict:
4692
4693
4694
4695
4696
4697
4698
4699
4700
                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
4701
4702
4703


class SequenceSummary(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
4704
4705
4706
4707
    r"""
    Compute a single vector summary of a sequence hidden states.

    Args:
4708
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
4709
4710
            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
4711

4712
            - **summary_type** (`str`) -- The method to use to make this summary. Accepted values are:
Sylvain Gugger's avatar
Sylvain Gugger committed
4713

4714
4715
4716
4717
4718
                - `"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
4719

4720
            - **summary_use_proj** (`bool`) -- Add a projection after the vector extraction.
Sylvain Gugger's avatar
Sylvain Gugger committed
4721
4722
4723
4724
4725
4726
            - **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
4727
    """
4728

4729
    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
4730
        super().__init__()
thomwolf's avatar
thomwolf committed
4731

4732
        self.summary_type = getattr(config, "summary_type", "last")
4733
        if self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
4734
4735
4736
4737
4738
            # 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
4739
        self.summary = Identity()
4740
4741
        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:
4742
                num_classes = config.num_labels
thomwolf's avatar
thomwolf committed
4743
4744
4745
4746
            else:
                num_classes = config.hidden_size
            self.summary = nn.Linear(config.hidden_size, num_classes)

4747
        activation_string = getattr(config, "summary_activation", None)
Lysandre's avatar
Lysandre committed
4748
        self.activation: Callable = get_activation(activation_string) if activation_string else Identity()
thomwolf's avatar
thomwolf committed
4749

thomwolf's avatar
thomwolf committed
4750
        self.first_dropout = Identity()
4751
        if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
4752
4753
            self.first_dropout = nn.Dropout(config.summary_first_dropout)

thomwolf's avatar
thomwolf committed
4754
        self.last_dropout = Identity()
4755
        if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
4756
            self.last_dropout = nn.Dropout(config.summary_last_dropout)
thomwolf's avatar
thomwolf committed
4757

Sylvain Gugger's avatar
Sylvain Gugger committed
4758
4759
4760
4761
4762
4763
4764
    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:
4765
            hidden_states (`torch.FloatTensor` of shape `[batch_size, seq_len, hidden_size]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
4766
                The hidden states of the last layer.
4767
            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
4768
                Used if `summary_type == "cls_index"` and takes the last token of the sequence as classification token.
Sylvain Gugger's avatar
Sylvain Gugger committed
4769
4770

        Returns:
4771
            `torch.FloatTensor`: The summary of the sequence hidden states.
thomwolf's avatar
thomwolf committed
4772
        """
4773
        if self.summary_type == "last":
thomwolf's avatar
thomwolf committed
4774
            output = hidden_states[:, -1]
4775
        elif self.summary_type == "first":
thomwolf's avatar
thomwolf committed
4776
            output = hidden_states[:, 0]
4777
        elif self.summary_type == "mean":
thomwolf's avatar
thomwolf committed
4778
            output = hidden_states.mean(dim=1)
4779
        elif self.summary_type == "cls_index":
thomwolf's avatar
thomwolf committed
4780
            if cls_index is None:
Lysandre's avatar
Lysandre committed
4781
4782
4783
4784
4785
                cls_index = torch.full_like(
                    hidden_states[..., :1, :],
                    hidden_states.shape[-2] - 1,
                    dtype=torch.long,
                )
thomwolf's avatar
thomwolf committed
4786
            else:
thomwolf's avatar
thomwolf committed
4787
                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
4788
                cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
thomwolf's avatar
thomwolf committed
4789
            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
4790
4791
            output = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, XX, hidden_size)
        elif self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
4792
4793
            raise NotImplementedError

4794
        output = self.first_dropout(output)
thomwolf's avatar
thomwolf committed
4795
4796
        output = self.summary(output)
        output = self.activation(output)
4797
        output = self.last_dropout(output)
thomwolf's avatar
thomwolf committed
4798
4799
4800
4801

        return output


4802
def unwrap_model(model: nn.Module) -> nn.Module:
4803
4804
4805
4806
    """
    Recursively unwraps a model from potential containers (as used in distributed training).

    Args:
4807
        model (`torch.nn.Module`): The model to unwrap.
4808
4809
4810
4811
4812
4813
    """
    # 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
4814
4815


4816
def expand_device_map(device_map, param_names, start_prefix):
Sylvain Gugger's avatar
Sylvain Gugger committed
4817
4818
4819
4820
    """
    Expand a device map to return the correspondance parameter name to device.
    """
    new_device_map = {}
4821
    param_names = [p[len(start_prefix) :] for p in param_names if p.startswith(start_prefix)]
Sylvain Gugger's avatar
Sylvain Gugger committed
4822
    for module, device in device_map.items():
4823
4824
4825
        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
4826
4827
4828
    return new_device_map


4829
def get_disk_only_shard_files(device_map, sharded_metadata, start_prefix):
Sylvain Gugger's avatar
Sylvain Gugger committed
4830
4831
4832
    """
    Returns the list of shard files containing only weights offloaded to disk.
    """
4833
4834
4835
4836

    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
4837
    files_content = collections.defaultdict(list)
4838
    for weight_name, filename in weight_map.items():
Sylvain Gugger's avatar
Sylvain Gugger committed
4839
4840
4841
4842
4843
        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"}]