modeling_utils.py 185 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 gc
Yih-Dar's avatar
Yih-Dar committed
18
import inspect
Sylvain Gugger's avatar
Sylvain Gugger committed
19
import json
20
import os
21
import re
Sylvain Gugger's avatar
Sylvain Gugger committed
22
23
import shutil
import tempfile
24
import warnings
25
from contextlib import contextmanager
26
from dataclasses import dataclass
27
from functools import partial
28
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
29
30

import torch
31
from packaging import version
Sylvain Gugger's avatar
Sylvain Gugger committed
32
from torch import Tensor, nn
33
from torch.nn import CrossEntropyLoss
34

35
from .activations import get_activation
36
from .configuration_utils import PretrainedConfig
37
from .deepspeed import deepspeed_config, is_deepspeed_zero3_enabled
38
from .dynamic_module_utils import custom_object_save
39
from .generation import GenerationConfig, GenerationMixin
40
41
42
43
from .pytorch_utils import (  # noqa: F401
    Conv1D,
    apply_chunking_to_forward,
    find_pruneable_heads_and_indices,
Thomas Wang's avatar
Thomas Wang committed
44
    id_tensor_storage,
45
46
47
48
    prune_conv1d_layer,
    prune_layer,
    prune_linear_layer,
)
49
from .utils import (
Aymeric Augustin's avatar
Aymeric Augustin committed
50
    DUMMY_INPUTS,
51
    FLAX_WEIGHTS_NAME,
52
53
    SAFE_WEIGHTS_INDEX_NAME,
    SAFE_WEIGHTS_NAME,
54
55
    TF2_WEIGHTS_NAME,
    TF_WEIGHTS_NAME,
Sylvain Gugger's avatar
Sylvain Gugger committed
56
    WEIGHTS_INDEX_NAME,
57
    WEIGHTS_NAME,
58
    ContextManagers,
59
    ModelOutput,
Sylvain Gugger's avatar
Sylvain Gugger committed
60
    PushToHubMixin,
61
    cached_file,
62
    copy_func,
63
    download_url,
64
    has_file,
65
    is_accelerate_available,
66
    is_bitsandbytes_available,
67
    is_offline_mode,
68
    is_optimum_available,
69
    is_remote_url,
70
    is_safetensors_available,
71
    is_torch_tpu_available,
72
    logging,
Sylvain Gugger's avatar
Sylvain Gugger committed
73
    replace_return_docstrings,
74
)
75
76
from .utils.hub import convert_file_size_to_int, get_checkpoint_shard_files
from .utils.import_utils import ENV_VARS_TRUE_VALUES, importlib_metadata, is_sagemaker_mp_enabled
77
from .utils.quantization_config import BitsAndBytesConfig
78
from .utils.versions import require_version_core
79

Aymeric Augustin's avatar
Aymeric Augustin committed
80

81
82
83
XLA_USE_BF16 = os.environ.get("XLA_USE_BF16", "0").upper()
XLA_DOWNCAST_BF16 = os.environ.get("XLA_DOWNCAST_BF16", "0").upper()

84
85
86
if is_accelerate_available():
    from accelerate import dispatch_model, infer_auto_device_map, init_empty_weights
    from accelerate.utils import (
87
        check_tied_parameters_on_same_device,
88
        find_tied_parameters,
89
        get_balanced_memory,
90
91
92
93
94
95
        load_offloaded_weights,
        offload_weight,
        save_offload_index,
        set_module_tensor_to_device,
    )

96
97
98
99
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
100

Lysandre Debut's avatar
Lysandre Debut committed
101
logger = logging.get_logger(__name__)
102

103
104
105
106

_init_weights = True


107
108
109
110
111
112
113
114
115
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


116
117
118
119
120
121
122
123
@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
124
    old_init_weights = _init_weights
125
126
127
128
129
    if _enable:
        _init_weights = False
    try:
        yield
    finally:
130
        _init_weights = old_init_weights
131
132


thomwolf's avatar
thomwolf committed
133
134
135
136
137
try:
    from torch.nn import Identity
except ImportError:
    # Older PyTorch compatibility
    class Identity(nn.Module):
Lysandre's avatar
Lysandre committed
138
        r"""A placeholder identity operator that is argument-insensitive."""
139

thomwolf's avatar
thomwolf committed
140
        def __init__(self, *args, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
141
            super().__init__()
thomwolf's avatar
thomwolf committed
142
143
144
145

        def forward(self, input):
            return input

146

Lysandre Debut's avatar
Lysandre Debut committed
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
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


162
163
164
165
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
166
167
168
    try:
        return next(parameter.parameters()).dtype
    except StopIteration:
Sylvain Gugger's avatar
Sylvain Gugger committed
169
        # For nn.DataParallel compatibility in PyTorch > 1.5
Lysandre Debut's avatar
Lysandre Debut committed
170
171
172
173
174
175
176
177
178
179

        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


180
181
def get_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
182
    Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.
183
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
184
185
186
187
    last_dtype = None
    for t in parameter.parameters():
        last_dtype = t.dtype
        if t.is_floating_point():
188
189
190
            # 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
191
192
193
194
195
            # 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:
196
                    return torch.bfloat16
197
198
                if t.dtype == torch.double:
                    return torch.float32
Sylvain Gugger's avatar
Sylvain Gugger committed
199
            return t.dtype
200

Sylvain Gugger's avatar
Sylvain Gugger committed
201
202
203
    if last_dtype is not None:
        # if no floating dtype was found return whatever the first dtype is
        return last_dtype
204

205
206
207
208
209
210
211
212
213
214
215
216
217
    # 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:
218
219
        # fallback to the last dtype
        return last_tuple[1].dtype
220

221
222
223
224
225
226
227
    # fallback to buffer dtype
    for t in parameter.buffers():
        last_dtype = t.dtype
        if t.is_floating_point():
            return t.dtype
    return last_dtype

228
229
230
231
232
233
234
235
236
237
238
239
240
241

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
242
    Returns the first found floating dtype in `state_dict` if there is one, otherwise returns the first dtype.
243
244
245
246
247
248
249
    """
    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
250
        return next(state_dict.values()).dtype
251
252


Sylvain Gugger's avatar
Sylvain Gugger committed
253
254
255
256
257
258
259
260
261
262
263
264
265
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
266
    bit_search = re.search(r"[^\d](\d+)$", str(dtype))
Sylvain Gugger's avatar
Sylvain Gugger committed
267
268
269
270
271
272
    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


273
274
275
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
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
    """
    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}>

    If one of the model's weight is bigger that `max_sahrd_size`, it will end up in its own sub-checkpoint which will
    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"`).
297
298
        weights_name (`str`, *optional*, defaults to `"pytorch_model.bin"`):
            The name of the model save file.
Sylvain Gugger's avatar
Sylvain Gugger committed
299
300
301
    """
    max_shard_size = convert_file_size_to_int(max_shard_size)

Thomas Wang's avatar
Thomas Wang committed
302
303
    sharded_state_dicts = [{}]
    last_block_size = 0
Sylvain Gugger's avatar
Sylvain Gugger committed
304
    total_size = 0
Thomas Wang's avatar
Thomas Wang committed
305
    storage_id_to_block = {}
Sylvain Gugger's avatar
Sylvain Gugger committed
306
307

    for key, weight in state_dict.items():
308
309
310
311
312
313
        # 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
314
315
316
317
318
319
320

        # 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
321
322
        weight_size = weight.numel() * dtype_byte_size(weight.dtype)

Sylvain Gugger's avatar
Sylvain Gugger committed
323
324
325
        # 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
326
327
            sharded_state_dicts.append({})
            last_block_size = 0
Sylvain Gugger's avatar
Sylvain Gugger committed
328

Thomas Wang's avatar
Thomas Wang committed
329
330
        sharded_state_dicts[-1][key] = weight
        last_block_size += weight_size
Sylvain Gugger's avatar
Sylvain Gugger committed
331
        total_size += weight_size
Thomas Wang's avatar
Thomas Wang committed
332
        storage_id_to_block[storage_id] = len(sharded_state_dicts) - 1
Sylvain Gugger's avatar
Sylvain Gugger committed
333
334
335

    # If we only have one shard, we return it
    if len(sharded_state_dicts) == 1:
336
        return {weights_name: sharded_state_dicts[0]}, None
Sylvain Gugger's avatar
Sylvain Gugger committed
337
338
339
340
341

    # Otherwise, let's build the index
    weight_map = {}
    shards = {}
    for idx, shard in enumerate(sharded_state_dicts):
342
343
344
345
        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
346
347
348
349
350
351
352
353
354
355
        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


356
def load_sharded_checkpoint(model, folder, strict=True, prefer_safe=True):
357
358
359
360
361
362
363
364
365
366
367
368
369
    """
    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.
370
371
372
        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.
373
374
375
376
377
378
379
380

    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)
381
    safe_index_file = os.path.join(folder, SAFE_WEIGHTS_INDEX_NAME)
382

383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
    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:
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
        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)

426
427
    loader = safe_load_file if load_safe else partial(torch.load, map_location="cpu")

428
    for shard_file in shard_files:
429
        state_dict = loader(os.path.join(folder, shard_file))
430
431
        model.load_state_dict(state_dict, strict=False)

432
        # Make sure memory is freed before we load the next state dict.
433
434
435
436
437
438
439
        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
440
441
442
443
def load_state_dict(checkpoint_file: Union[str, os.PathLike]):
    """
    Reads a PyTorch checkpoint file, returning properly formatted errors if they arise.
    """
444
445
446
447
448
449
450
451
452
453
454
455
456
457
    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."
            )
        elif metadata["format"] != "pt":
            raise NotImplementedError(
                f"Conversion from a {metadata['format']} safetensors archive to PyTorch is not implemented yet."
            )
        return safe_load_file(checkpoint_file)
Sylvain Gugger's avatar
Sylvain Gugger committed
458
459
460
461
462
    try:
        return torch.load(checkpoint_file, map_location="cpu")
    except Exception as e:
        try:
            with open(checkpoint_file) as f:
463
                if f.read(7) == "version":
Sylvain Gugger's avatar
Sylvain Gugger committed
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
                    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."
            )


482
483
484
485
486
487
488
489
490
491
492
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
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
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.
519
    def load(module: nn.Module, state_dict, prefix=""):
Sylvain Gugger's avatar
Sylvain Gugger committed
520
521
        local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
        args = (state_dict, prefix, local_metadata, True, [], [], error_msgs)
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
        # 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
541
542
543

        for name, child in module._modules.items():
            if child is not None:
544
                load(child, state_dict, prefix + name + ".")
Sylvain Gugger's avatar
Sylvain Gugger committed
545

546
547
548
549
    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
550
551
552
553

    return error_msgs


554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
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)


603
604
605
606
607
608
609
610
611
612
613
614
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,
615
    is_quantized=False,
Sylvain Gugger's avatar
Sylvain Gugger committed
616
    is_safetensors=False,
617
    keep_in_fp32_modules=None,
618
):
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
    """
    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.

636
637
    if is_quantized:
        from .utils.bitsandbytes import set_module_quantized_tensor_to_device
638

639
640
    error_msgs = []

641
642
643
644
645
646
647
648
649
650
651
652
653
    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)
654

655
656
657
658
659
660
661
662
663
    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
664
        set_module_kwargs = {}
665

666
        # We convert floating dtypes to the `dtype` passed. We want to keep the buffers/params
667
668
        # in int/uint/bool and not cast them.
        if dtype is not None and torch.is_floating_point(param):
669
670
671
672
673
674
            if (
                keep_in_fp32_modules is not None
                and any(module_to_keep_in_fp32 in param_name for module_to_keep_in_fp32 in keep_in_fp32_modules)
                and dtype == torch.float16
            ):
                param = param.to(torch.float32)
675
676
677
678
679

                # 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
680
681
            else:
                param = param.to(dtype)
682
683
684
685
686
687
688
689
690
691
692
693

        # 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)
694

695
696
        set_module_kwargs["value"] = param

697
698
699
700
701
702
703
704
705
706
707
        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]
708

709
        if param_device == "disk":
Sylvain Gugger's avatar
Sylvain Gugger committed
710
711
            if not is_safetensors:
                offload_index = offload_weight(param, param_name, offload_folder, offload_index)
712
713
        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)
714
        elif not is_quantized:
715
716
            # For backward compatibility with older versions of `accelerate`
            set_module_tensor_to_device(model, param_name, param_device, **set_module_kwargs)
717
        else:
718
719
720
721
722
723
            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:
724
                set_module_quantized_tensor_to_device(
725
726
                    model, param_name, param_device, value=param, fp16_statistics=fp16_statistics
                )
727
728

    return error_msgs, offload_index, state_dict_index
729
730


731
732
733
734
735
736
737
738
739
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


740
class ModuleUtilsMixin:
Julien Chaumond's avatar
Julien Chaumond committed
741
    """
742
    A few utilities for `torch.nn.Modules`, to be used as a mixin.
Julien Chaumond's avatar
Julien Chaumond committed
743
744
    """

745
746
747
748
    @staticmethod
    def _hook_rss_memory_pre_forward(module, *args, **kwargs):
        try:
            import psutil
Sylvain Gugger's avatar
Sylvain Gugger committed
749
        except ImportError:
750
751
752
753
754
755
756
757
758
759
760
            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
761
        except ImportError:
762
763
764
765
766
767
768
769
770
771
            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
772
773
774
        """
        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
775
776
        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()`.
777
778
779
780
781
782
783
        """
        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
784
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
785
        Reset the `mem_rss_diff` attribute of each module (see [`~modeling_utils.ModuleUtilsMixin.add_memory_hooks`]).
Sylvain Gugger's avatar
Sylvain Gugger committed
786
        """
787
788
789
790
791
        for module in self.modules():
            module.mem_rss_diff = 0
            module.mem_rss_post_forward = 0
            module.mem_rss_pre_forward = 0

792
    @property
Sylvain Gugger's avatar
Sylvain Gugger committed
793
    def device(self) -> torch.device:
794
        """
795
        `torch.device`: The device on which the module is (assuming that all the module parameters are on the same
796
        device).
797
        """
Lysandre Debut's avatar
Lysandre Debut committed
798
        return get_parameter_device(self)
799

800
    @property
801
    def dtype(self) -> torch.dtype:
802
        """
803
        `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).
804
        """
Lysandre Debut's avatar
Lysandre Debut committed
805
        return get_parameter_dtype(self)
806
807

    def invert_attention_mask(self, encoder_attention_mask: Tensor) -> Tensor:
Sylvain Gugger's avatar
Sylvain Gugger committed
808
809
810
811
        """
        Invert an attention mask (e.g., switches 0. and 1.).

        Args:
812
            encoder_attention_mask (`torch.Tensor`): An attention mask.
Sylvain Gugger's avatar
Sylvain Gugger committed
813
814

        Returns:
815
            `torch.Tensor`: The inverted attention mask.
Sylvain Gugger's avatar
Sylvain Gugger committed
816
        """
817
818
819
820
821
822
823
824
825
826
        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
827
        encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * torch.finfo(self.dtype).min
828

829
830
        return encoder_extended_attention_mask

831
    @staticmethod
832
833
834
835
836
837
838
    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
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
        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

859
    def get_extended_attention_mask(
860
        self, attention_mask: Tensor, input_shape: Tuple[int], device: torch.device = None, dtype: torch.float = None
861
    ) -> Tensor:
Sylvain Gugger's avatar
Sylvain Gugger committed
862
863
        """
        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
864
865

        Arguments:
866
            attention_mask (`torch.Tensor`):
Sylvain Gugger's avatar
Sylvain Gugger committed
867
                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
868
            input_shape (`Tuple[int]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
869
                The shape of the input to the model.
870
871

        Returns:
872
            `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
873
        """
Yih-Dar's avatar
Yih-Dar committed
874
875
876
        if dtype is None:
            dtype = self.dtype

877
878
879
880
881
882
        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
                )
883
884
885
886
887
888
889
890
891
        # 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:
892
                extended_attention_mask = ModuleUtilsMixin.create_extended_attention_mask_for_decoder(
893
894
                    input_shape, attention_mask, device
                )
895
896
897
898
            else:
                extended_attention_mask = attention_mask[:, None, None, :]
        else:
            raise ValueError(
899
                f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"
900
901
902
903
            )

        # 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
904
        # positions we want to attend and the dtype's smallest value for masked positions.
905
906
        # 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
907
908
        extended_attention_mask = extended_attention_mask.to(dtype=dtype)  # fp16 compatibility
        extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(dtype).min
909
910
        return extended_attention_mask

Sylvain Gugger's avatar
Sylvain Gugger committed
911
912
913
    def get_head_mask(
        self, head_mask: Optional[Tensor], num_hidden_layers: int, is_attention_chunked: bool = False
    ) -> Tensor:
914
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
915
916
917
        Prepare the head mask if needed.

        Args:
918
            head_mask (`torch.Tensor` with shape `[num_heads]` or `[num_hidden_layers x num_heads]`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
919
                The mask indicating if we should keep the heads or not (1.0 for keep, 0.0 for discard).
920
            num_hidden_layers (`int`):
Sylvain Gugger's avatar
Sylvain Gugger committed
921
                The number of hidden layers in the model.
922
            is_attention_chunked (`bool`, *optional*, defaults to `False`):
Sylvain Gugger's avatar
Sylvain Gugger committed
923
924
                Whether or not the attentions scores are computed by chunks or not.

925
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
926
927
            `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.
928
929
930
        """
        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
931
932
            if is_attention_chunked is True:
                head_mask = head_mask.unsqueeze(-1)
933
934
935
936
937
938
939
940
941
942
943
944
945
        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()}"
946
        head_mask = head_mask.to(dtype=self.dtype)  # switch to float if need + fp16 compatibility
947
948
        return head_mask

949
950
951
952
953
    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:
954
            only_trainable (`bool`, *optional*, defaults to `False`):
955
956
                Whether or not to return only the number of trainable parameters

957
            exclude_embeddings (`bool`, *optional*, defaults to `False`):
958
959
960
                Whether or not to return only the number of non-embeddings parameters

        Returns:
961
            `int`: The number of parameters.
962
963
        """

964
965
966
967
968
969
970
971
972
973
        if exclude_embeddings:
            embedding_param_names = [
                f"{name}.weight" for name, module_type in self.named_modules() if isinstance(module_type, nn.Embedding)
            ]
            non_embedding_parameters = [
                parameter for name, parameter in self.named_parameters() if name not in embedding_param_names
            ]
            return sum(p.numel() for p in non_embedding_parameters if p.requires_grad or not only_trainable)
        else:
            return sum(p.numel() for p in self.parameters() if p.requires_grad or not only_trainable)
974
975
976
977
978
979

    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:
980
            inputs (`dict`): The model inputs.
981
982

        Returns:
983
            `int`: The total number of tokens.
984
        """
985
986
        if not hasattr(self, "warnings_issued"):
            self.warnings_issued = {}
987
988
        if self.main_input_name in input_dict:
            return input_dict[self.main_input_name].numel()
989
        elif "estimate_tokens" not in self.warnings_issued:
990
            logger.warning(
991
992
                "Could not estimate the number of tokens of the input, floating-point operations will not be computed"
            )
993
994
            self.warnings_issued["estimate_tokens"] = True
        return 0
995
996
997
998
999
1000
1001

    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
1002
1003
        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
1004
        re-use e.g. Albert or Universal Transformers, or if doing long-range modeling with very high sequence lengths.
1005
1006

        Args:
1007
            batch_size (`int`):
1008
1009
                The batch size for the forward pass.

1010
            sequence_length (`int`):
1011
1012
                The number of tokens in each line of the batch.

1013
            exclude_embeddings (`bool`, *optional*, defaults to `True`):
1014
1015
1016
                Whether or not to count embedding and softmax operations.

        Returns:
1017
            `int`: The number of floating-point operations.
1018
1019
1020
1021
        """

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

Julien Chaumond's avatar
Julien Chaumond committed
1022

Sylvain Gugger's avatar
Sylvain Gugger committed
1023
class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin):
1024
1025
    r"""
    Base class for all models.
1026

Sylvain Gugger's avatar
Sylvain Gugger committed
1027
1028
    [`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:
1029

1030
1031
        - resize the input embeddings,
        - prune heads in the self-attention heads.
1032

1033
    Class attributes (overridden by derived classes):
Sylvain Gugger's avatar
Sylvain Gugger committed
1034

Sylvain Gugger's avatar
Sylvain Gugger committed
1035
1036
1037
1038
        - **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:
1039

Sylvain Gugger's avatar
Sylvain Gugger committed
1040
1041
            - **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.
1042
            - **path** (`str`) -- A path to the TensorFlow checkpoint.
1043

Sylvain Gugger's avatar
Sylvain Gugger committed
1044
1045
        - **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.
1046
        - **is_parallelizable** (`bool`) -- A flag indicating whether this model supports model parallelization.
Sylvain Gugger's avatar
Sylvain Gugger committed
1047
1048
        - **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).
1049
    """
1050
    config_class = None
1051
    base_model_prefix = ""
1052
    main_input_name = "input_ids"
1053
    _auto_class = None
1054
    _no_split_modules = None
1055
    _skip_keys_device_placement = None
1056
    _keep_in_fp32_modules = None
1057

1058
1059
    # 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.
1060
    _keys_to_ignore_on_load_missing = None
1061
1062
1063
    # 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.
1064
    _keys_to_ignore_on_load_unexpected = None
1065
1066
    # 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)
1067
    _keys_to_ignore_on_save = None
Sylvain Gugger's avatar
Sylvain Gugger committed
1068
1069
    # a list of `state_dict` keys that are potentially tied to another key in the state_dict.
    _tied_weights_keys = None
1070

1071
    is_parallelizable = False
1072
    supports_gradient_checkpointing = False
1073

1074
    @property
1075
    def dummy_inputs(self) -> Dict[str, torch.Tensor]:
1076
        """
1077
        `Dict[str, torch.Tensor]`: Dummy inputs to do a forward pass in the network.
1078
        """
1079
        return {"input_ids": torch.tensor(DUMMY_INPUTS)}
1080

1081
1082
1083
1084
1085
1086
1087
    @property
    def framework(self) -> str:
        """
        :str: Identifies that this is a PyTorch model.
        """
        return "pt"

1088
    def __init__(self, config: PretrainedConfig, *inputs, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
1089
        super().__init__()
1090
1091
        if not isinstance(config, PretrainedConfig):
            raise ValueError(
1092
1093
1094
                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)`"
1095
            )
1096
        # Save config and origin of the pretrained weights if given in model
1097
        self.config = config
1098
        self.name_or_path = config.name_or_path
1099
        self.warnings_issued = {}
1100
        self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114

    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")
1115

1116
1117
1118
1119
1120
1121
    @classmethod
    def _from_config(cls, config, **kwargs):
        """
        All context managers that the model should be initialized under go here.

        Args:
1122
1123
            torch_dtype (`torch.dtype`, *optional*):
                Override the default `torch.dtype` and load the model under this dtype.
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
        """
        torch_dtype = kwargs.pop("torch_dtype", None)

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

        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
1138
            with deepspeed.zero.Init(config_dict_or_path=deepspeed_config()):
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
                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

    @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:
1156
            dtype (`torch.dtype`):
1157
1158
1159
                a floating dtype to set to.

        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
1160
1161
            `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`.
1162

1163
1164
        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.
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
        """
        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

1176
    @property
1177
1178
    def base_model(self) -> nn.Module:
        """
1179
        `torch.nn.Module`: The main body of the model.
1180
        """
1181
        return getattr(self, self.base_model_prefix, self)
thomwolf's avatar
thomwolf committed
1182

1183
1184
    @classmethod
    def can_generate(cls) -> bool:
1185
1186
1187
1188
1189
1190
1191
        """
        Returns whether this model can generate sequences with `.generate()`.

        Returns:
            `bool`: Whether this model can generate sequences with `.generate()`.
        """
        # Detects whether `prepare_inputs_for_generation` has been overwritten, which is a requirement for generation
1192
        if "GenerationMixin" in str(cls.prepare_inputs_for_generation):
1193
1194
1195
            return False
        return True

1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
    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()

1213
    def get_input_embeddings(self) -> nn.Module:
1214
1215
1216
1217
        """
        Returns the model's input embeddings.

        Returns:
1218
            `nn.Module`: A torch module mapping vocabulary to hidden states.
thomwolf's avatar
thomwolf committed
1219
        """
1220
        base_model = getattr(self, self.base_model_prefix, self)
thomwolf's avatar
thomwolf committed
1221
1222
1223
1224
        if base_model is not self:
            return base_model.get_input_embeddings()
        else:
            raise NotImplementedError
thomwolf's avatar
thomwolf committed
1225

1226
    def set_input_embeddings(self, value: nn.Module):
1227
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
1228
        Set model's input embeddings.
1229
1230

        Args:
1231
            value (`nn.Module`): A module mapping vocabulary to hidden states.
thomwolf's avatar
thomwolf committed
1232
1233
1234
1235
1236
1237
        """
        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
1238

1239
    def get_output_embeddings(self) -> nn.Module:
1240
1241
1242
1243
        """
        Returns the model's output embeddings.

        Returns:
1244
            `nn.Module`: A torch module mapping hidden states to vocabulary.
thomwolf's avatar
thomwolf committed
1245
        """
1246
        return None  # Overwrite for models with output embeddings
thomwolf's avatar
thomwolf committed
1247

1248
1249
1250
1251
    def _init_weights(self, module):
        """
        Initialize the weights. This method should be overridden by derived class.
        """
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
        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
1262

1263
    def tie_weights(self):
1264
1265
        """
        Tie the weights between the input embeddings and the output embeddings.
1266

Sylvain Gugger's avatar
Sylvain Gugger committed
1267
1268
        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
1269
        """
1270
1271
1272
1273
        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
1274

1275
        if getattr(self.config, "is_encoder_decoder", False) and getattr(self.config, "tie_encoder_decoder", False):
Weizhen's avatar
Weizhen committed
1276
1277
            if hasattr(self, self.base_model_prefix):
                self = getattr(self, self.base_model_prefix)
1278
1279
            self._tie_encoder_decoder_weights(self.encoder, self.decoder, self.base_model_prefix)

Sylvain Gugger's avatar
Sylvain Gugger committed
1280
1281
1282
1283
        for module in self.modules():
            if hasattr(module, "_tie_weights"):
                module._tie_weights()

1284
1285
1286
    @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
1287
1288
        if decoder.__class__ != encoder.__class__:
            logger.info(
Sylvain Gugger's avatar
Sylvain Gugger committed
1289
1290
                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
1291
            )
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301

        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
1302
            ), f"{decoder_pointer} and {encoder_pointer} have to be of type nn.Module"
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
            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}"

1318
                all_encoder_weights = {module_name + "/" + sub_name for sub_name in encoder_modules.keys()}
1319
1320
1321
1322
1323
                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
1324
1325
1326
                        if not isinstance(decoder_modules[decoder_name], type(encoder_modules[encoder_name])) and len(
                            encoder_modules
                        ) != len(decoder_modules):
1327
1328
                            # 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
1329
                            # thus skip this step and subtract one layer pos from encoder
1330
1331
1332
1333
1334
1335
                            encoder_layer_pos -= 1
                            continue
                    elif name not in encoder_modules:
                        continue
                    elif depth > 500:
                        raise ValueError(
Sylvain Gugger's avatar
Sylvain Gugger committed
1336
1337
                            "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."
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
                        )
                    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}"
            )

1359
    def _tie_or_clone_weights(self, output_embeddings, input_embeddings):
Lysandre's avatar
Lysandre committed
1360
        """Tie or clone module weights depending of whether we are using TorchScript or not"""
thomwolf's avatar
thomwolf committed
1361
        if self.config.torchscript:
1362
            output_embeddings.weight = nn.Parameter(input_embeddings.weight.clone())
thomwolf's avatar
thomwolf committed
1363
        else:
1364
            output_embeddings.weight = input_embeddings.weight
thomwolf's avatar
thomwolf committed
1365

Sam Shleifer's avatar
Sam Shleifer committed
1366
        if getattr(output_embeddings, "bias", None) is not None:
1367
            output_embeddings.bias.data = nn.functional.pad(
1368
                output_embeddings.bias.data,
Lysandre's avatar
Lysandre committed
1369
1370
1371
1372
                (
                    0,
                    output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0],
                ),
1373
1374
                "constant",
                0,
1375
            )
1376
        if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
1377
            output_embeddings.out_features = input_embeddings.num_embeddings
1378

1379
    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None) -> nn.Embedding:
1380
        """
1381
        Resizes input token embeddings matrix of the model if `new_num_tokens != config.vocab_size`.
1382

1383
        Takes care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
thomwolf's avatar
thomwolf committed
1384

1385
        Arguments:
1386
            new_num_tokens (`int`, *optional*):
1387
                The number of new tokens in the embedding matrix. Increasing the size will add newly initialized
Sylvain Gugger's avatar
Sylvain Gugger committed
1388
1389
                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.
1390
1391

        Return:
1392
            `torch.nn.Embedding`: Pointer to the input tokens Embeddings Module of the model.
thomwolf's avatar
thomwolf committed
1393
        """
1394
        model_embeds = self._resize_token_embeddings(new_num_tokens)
thomwolf's avatar
thomwolf committed
1395
1396
        if new_num_tokens is None:
            return model_embeds
thomwolf's avatar
thomwolf committed
1397
1398
1399

        # Update base model and current model config
        self.config.vocab_size = new_num_tokens
1400
        self.vocab_size = new_num_tokens
thomwolf's avatar
thomwolf committed
1401
1402

        # Tie weights again if needed
1403
        self.tie_weights()
thomwolf's avatar
thomwolf committed
1404

thomwolf's avatar
thomwolf committed
1405
1406
        return model_embeds

1407
    def _resize_token_embeddings(self, new_num_tokens):
thomwolf's avatar
thomwolf committed
1408
1409
1410
        old_embeddings = self.get_input_embeddings()
        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
        self.set_input_embeddings(new_embeddings)
1411
1412
1413
1414
1415
1416
1417

        # 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()
            new_lm_head = self._get_resized_lm_head(old_lm_head, new_num_tokens)
            self.set_output_embeddings(new_lm_head)

thomwolf's avatar
thomwolf committed
1418
        return self.get_input_embeddings()
1419

1420
    def _get_resized_embeddings(
1421
1422
        self, old_embeddings: nn.Embedding, new_num_tokens: Optional[int] = None
    ) -> nn.Embedding:
1423
1424
1425
        """
        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
1426
1427

        Args:
1428
            old_embeddings (`torch.nn.Embedding`):
1429
                Old embeddings to be resized.
1430
            new_num_tokens (`int`, *optional*):
1431
                New number of tokens in the embedding matrix.
1432
1433

                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
1434
                vectors from the end. If not provided or `None`, just returns a pointer to the input tokens
1435
                `torch.nn.Embedding` module of the model without doing anything.
1436
1437

        Return:
1438
1439
            `torch.nn.Embedding`: Pointer to the resized Embedding Module or the old Embedding Module if
            `new_num_tokens` is `None`
1440
1441
1442
1443
        """
        if new_num_tokens is None:
            return old_embeddings

1444
1445
1446
1447
1448
1449
1450
1451
        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()

1452
1453
1454
        if old_num_tokens == new_num_tokens:
            return old_embeddings

1455
1456
        if not isinstance(old_embeddings, nn.Embedding):
            raise TypeError(
Sylvain Gugger's avatar
Sylvain Gugger committed
1457
1458
1459
                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}."
1460
1461
            )

1462
        # Build new embeddings
1463
        new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim)
1464
        new_embeddings.to(old_embeddings.weight.device, dtype=old_embeddings.weight.dtype)
1465
1466
1467
1468

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

1469
        # Copy token embeddings from the previous weights
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480

        # numbers of tokens to copy
        n = min(old_num_tokens, new_num_tokens)
        if is_deepspeed_zero3_enabled():
            import deepspeed

            with deepspeed.zero.GatheredParameters(old_embeddings.weight, modifier_rank=0):
                if torch.distributed.get_rank() == 0:
                    new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
        else:
            new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
1481
1482
1483

        return new_embeddings

1484
    def _get_resized_lm_head(
1485
1486
        self, old_lm_head: nn.Linear, new_num_tokens: Optional[int] = None, transposed: Optional[bool] = False
    ) -> nn.Linear:
1487
1488
1489
1490
1491
        """
        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:
1492
            old_lm_head (`torch.nn.Linear`):
1493
                Old lm head liner layer to be resized.
1494
            new_num_tokens (`int`, *optional*):
1495
1496
1497
                New number of tokens in the linear matrix.

                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
1498
                vectors from the end. If not provided or `None`, just returns a pointer to the input tokens
1499
1500
1501
                `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`.
1502
1503

        Return:
Sylvain Gugger's avatar
Sylvain Gugger committed
1504
1505
            `torch.nn.Linear`: Pointer to the resized Linear Module or the old Linear Module if `new_num_tokens` is
            `None`
1506
1507
1508
1509
        """
        if new_num_tokens is None:
            return old_lm_head

1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
        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()
            )
1521
1522
1523
1524
1525
1526

        if old_num_tokens == new_num_tokens:
            return old_lm_head

        if not isinstance(old_lm_head, nn.Linear):
            raise TypeError(
Sylvain Gugger's avatar
Sylvain Gugger committed
1527
1528
1529
                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}."
1530
1531
1532
1533
1534
            )

        # 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
1535
        new_lm_head = nn.Linear(*new_lm_head_shape, bias=has_new_lm_head_bias)
1536
        new_lm_head = new_lm_head.to(old_lm_head.weight.device, dtype=old_lm_head.weight.dtype)
1537
1538
1539
1540
1541
1542

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

        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)

1543
1544
1545
1546
        # XXX: put the long block of code in a wrapper
        if is_deepspeed_zero3_enabled():
            import deepspeed

1547
1548
            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):
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
                if torch.distributed.get_rank() == 0:
                    # 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]
1563
        else:
1564
1565
1566
1567
1568
            # 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]
1569

1570
1571
1572
            # 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]
1573
1574
1575

        return new_lm_head

1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
    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`"
        )

1588
    def init_weights(self):
1589
        """
1590
1591
        If needed prunes and maybe initializes weights. If using a custom `PreTrainedModel`, you need to implement any
        initialization logic in `_init_weights`.
1592
        """
1593
1594
1595
1596
        # Prune heads if needed
        if self.config.pruned_heads:
            self.prune_heads(self.config.pruned_heads)

1597
1598
        if _init_weights:
            # Initialize weights
1599
            self.apply(self._initialize_weights)
1600
1601
1602
1603

            # Tie weights should be skipped when not initializing all weights
            # since from_pretrained(...) calls tie weights anyways
            self.tie_weights()
1604

1605
1606
1607
    def prune_heads(self, heads_to_prune: Dict[int, List[int]]):
        """
        Prunes heads of the base model.
1608

1609
        Arguments:
1610
            heads_to_prune (`Dict[int, List[int]]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1611
1612
1613
                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
1614
        """
1615
        # save new sets of pruned heads as union of previously stored pruned heads and newly pruned heads
1616
        for layer, heads in heads_to_prune.items():
1617
1618
1619
            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

1620
        self.base_model._prune_heads(heads_to_prune)
thomwolf's avatar
thomwolf committed
1621

1622
    def gradient_checkpointing_enable(self):
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
        """
        Activates gradient checkpointing for the current model.

        Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint
        activations".
        """
        if not self.supports_gradient_checkpointing:
            raise ValueError(f"{self.__class__.__name__} does not support gradient checkpointing.")
        self.apply(partial(self._set_gradient_checkpointing, value=True))

1633
    def gradient_checkpointing_disable(self):
1634
1635
1636
1637
1638
1639
1640
1641
1642
        """
        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:
            self.apply(partial(self._set_gradient_checkpointing, value=False))

1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
    @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())

1653
1654
1655
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
1656
        is_main_process: bool = True,
1657
1658
        state_dict: Optional[dict] = None,
        save_function: Callable = torch.save,
Sylvain Gugger's avatar
Sylvain Gugger committed
1659
        push_to_hub: bool = False,
Sylvain Gugger's avatar
Sylvain Gugger committed
1660
        max_shard_size: Union[int, str] = "10GB",
1661
        safe_serialization: bool = False,
1662
        variant: Optional[str] = None,
Sylvain Gugger's avatar
Sylvain Gugger committed
1663
        **kwargs,
1664
    ):
1665
1666
        """
        Save a model and its configuration file to a directory, so that it can be re-loaded using the
1667
        [`~PreTrainedModel.from_pretrained`] class method.
1668

1669
        Arguments:
1670
            save_directory (`str` or `os.PathLike`):
1671
                Directory to which to save. Will be created if it doesn't exist.
1672
1673
1674
1675
            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.
1676
            state_dict (nested dictionary of `torch.Tensor`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1677
1678
1679
                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).
1680
            save_function (`Callable`):
1681
                The function to use to save the state dictionary. Useful on distributed training like TPUs when one
1682
1683
                need to replace `torch.save` by another method.
            push_to_hub (`bool`, *optional*, defaults to `False`):
1684
1685
1686
                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).
Sylvain Gugger's avatar
Sylvain Gugger committed
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
            max_shard_size (`int` or `str`, *optional*, defaults to `"10GB"`):
                The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size
                lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5MB"`).

                <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>

1698
1699
            safe_serialization (`bool`, *optional*, defaults to `False`):
                Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
1700
1701
            variant (`str`, *optional*):
                If specified, weights are saved in the format pytorch_model.<variant>.bin.
1702

Sylvain Gugger's avatar
Sylvain Gugger committed
1703
            kwargs:
1704
                Additional key word arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
1705
        """
1706
        # Checks if the model has been loaded in 8-bit
1707
        if getattr(self, "is_loaded_in_8bit", False) and getattr(self, "is_8bit_serializable", False):
1708
1709
            warnings.warn(
                "You are calling `save_pretrained` to a 8-bit converted model you may likely encounter unexepected"
1710
                " behaviors. If you want to save 8-bit models, make sure to have `bitsandbytes>0.37.2` installed.",
1711
1712
1713
                UserWarning,
            )

1714
1715
1716
1717
1718
        if getattr(self, "is_loaded_in_4bit", False):
            raise NotImplementedError(
                "You are calling `save_pretrained` on a 4-bit converted model. This is currently not supported"
            )

1719
1720
1721
1722
1723
        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")
1724
1725
        if safe_serialization and not is_safetensors_available():
            raise ImportError("`safe_serialization` requires the `safetensors library: `pip install safetensors`.")
1726

1727
        if os.path.isfile(save_directory):
1728
            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
1729
            return
1730

1731
1732
        os.makedirs(save_directory, exist_ok=True)

1733
1734
        if push_to_hub:
            commit_message = kwargs.pop("commit_message", None)
1735
            repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
1736
            repo_id = self._create_repo(repo_id, **kwargs)
1737
            files_timestamps = self._get_files_timestamps(save_directory)
1738

Julien Chaumond's avatar
Julien Chaumond committed
1739
        # Only save the model itself if we are using distributed training
1740
        model_to_save = unwrap_model(self)
1741

1742
1743
1744
1745
1746
        # 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
1747
1748
1749
        # Attach architecture to the config
        model_to_save.config.architectures = [model_to_save.__class__.__name__]

1750
1751
1752
1753
1754
        # 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)

1755
        # Save the config
1756
        if is_main_process:
1757
            model_to_save.config.save_pretrained(save_directory)
1758
1759
            if self.can_generate():
                model_to_save.generation_config.save_pretrained(save_directory)
1760
1761
1762
1763

        # Save the model
        if state_dict is None:
            state_dict = model_to_save.state_dict()
1764

1765
1766
1767
1768
1769
        # 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)

1770
        # Handle the case where some state_dict keys shouldn't be saved
1771
        if self._keys_to_ignore_on_save is not None:
1772
            for ignore_key in self._keys_to_ignore_on_save:
1773
1774
                if ignore_key in state_dict.keys():
                    del state_dict[ignore_key]
1775
1776
1777
1778
1779
        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():
Sylvain Gugger's avatar
Sylvain Gugger committed
1780
                ptrs[id_tensor_storage(tensor)].append(name)
1781
1782
1783
1784
1785
1786
1787

            # 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
1788
                if self._tied_weights_keys is not None:
1789
1790
                    found = 0
                    for name in sorted(names):
Sylvain Gugger's avatar
Sylvain Gugger committed
1791
                        matches_pattern = any(re.search(pat, name) for pat in self._tied_weights_keys)
1792
                        if matches_pattern and name in state_dict:
1793
1794
1795
                            found += 1
                            if found < len(names):
                                del state_dict[name]
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812

                # 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",
                )
1813

Sylvain Gugger's avatar
Sylvain Gugger committed
1814
        # Shard the model if it is too big.
1815
        weights_name = SAFE_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
1816
1817
        weights_name = _add_variant(weights_name, variant)

1818
        shards, index = shard_checkpoint(state_dict, max_shard_size=max_shard_size, weights_name=weights_name)
Sylvain Gugger's avatar
Sylvain Gugger committed
1819
1820
1821
1822

        # Clean the folder from a previous save
        for filename in os.listdir(save_directory):
            full_filename = os.path.join(save_directory, filename)
1823
1824
            # 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.
1825
            weights_no_suffix = weights_name.replace(".bin", "").replace(".safetensors", "")
1826
1827
1828

            # 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", "")
1829
            reg = re.compile(r"(.*?)-\d{5}-of-\d{5}")
1830

1831
            if (
1832
                filename.startswith(weights_no_suffix)
1833
1834
1835
                and os.path.isfile(full_filename)
                and filename not in shards.keys()
                and is_main_process
1836
                and reg.fullmatch(filename_no_suffix) is not None
1837
            ):
Sylvain Gugger's avatar
Sylvain Gugger committed
1838
                os.remove(full_filename)
1839

Sylvain Gugger's avatar
Sylvain Gugger committed
1840
1841
        # Save the model
        for shard_file, shard in shards.items():
1842
1843
1844
1845
1846
1847
            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
1848
1849

        if index is None:
1850
1851
            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
1852
        else:
1853
            save_index_file = SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
1854
            save_index_file = os.path.join(save_directory, _add_variant(save_index_file, variant))
Sylvain Gugger's avatar
Sylvain Gugger committed
1855
1856
1857
1858
1859
1860
1861
1862
1863
            # 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}."
            )
1864

Sylvain Gugger's avatar
Sylvain Gugger committed
1865
        if push_to_hub:
1866
            self._upload_modified_files(
1867
1868
1869
1870
1871
                save_directory,
                repo_id,
                files_timestamps,
                commit_message=commit_message,
                token=kwargs.get("use_auth_token"),
1872
            )
Sylvain Gugger's avatar
Sylvain Gugger committed
1873

1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
    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

1892
1893
    def to(self, *args, **kwargs):
        # Checks if the model has been loaded in 8-bit
1894
        if getattr(self, "is_quantized", False):
1895
            raise ValueError(
1896
                "`.to` is not supported for `4-bit` or `8-bit` models. Please use the model as it is, since the"
1897
1898
1899
1900
1901
1902
1903
                " model has already been set to the correct devices and casted to the correct `dtype`."
            )
        else:
            return super().to(*args, **kwargs)

    def half(self, *args):
        # Checks if the model has been loaded in 8-bit
1904
        if getattr(self, "is_quantized", False):
1905
            raise ValueError(
1906
                "`.half()` is not supported for `4-bit` or `8-bit` models. Please use the model as it is, since the"
1907
1908
1909
1910
1911
1912
1913
                " model has already been casted to the correct `dtype`."
            )
        else:
            return super().half(*args)

    def float(self, *args):
        # Checks if the model has been loaded in 8-bit
1914
        if getattr(self, "is_quantized", False):
1915
            raise ValueError(
1916
                "`.float()` is not supported for `4-bit` or `8-bit` models. Please use the model as it is, since the"
1917
1918
1919
1920
1921
                " model has already been casted to the correct `dtype`."
            )
        else:
            return super().float(*args)

1922
    @classmethod
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
    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,
    ):
1937
1938
        r"""
        Instantiate a pretrained pytorch model from a pre-trained model configuration.
1939

Sylvain Gugger's avatar
Sylvain Gugger committed
1940
1941
        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()`.
1942

1943
        The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
1944
1945
        pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
        task.
1946

1947
        The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
1948
        weights are discarded.
1949

1950
        Parameters:
1951
            pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
1952
1953
                Can be either:

1954
                    - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
Sylvain Gugger's avatar
Sylvain Gugger committed
1955
1956
                      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`.
1957
1958
1959
                    - 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
1960
1961
1962
                      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.
1963
                    - 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
1964
1965
                      `./flax_model/` containing `flax_model.msgpack`). In this case, `from_flax` should be set to
                      `True`.
1966
1967
1968
1969
1970
                    - `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*):
1971
1972
                Can be either:

1973
1974
                    - an instance of a class derived from [`PretrainedConfig`],
                    - a string or path valid as input to [`~PretrainedConfig.from_pretrained`].
1975

1976
                Configuration for the model to use instead of an automatically loaded configuration. Configuration can
1977
1978
                be automatically loaded when:

1979
                    - The model is a model provided by the library (loaded with the *model id* string of a pretrained
1980
                      model).
Sylvain Gugger's avatar
Sylvain Gugger committed
1981
1982
                    - The model was saved using [`~PreTrainedModel.save_pretrained`] and is reloaded by supplying the
                      save directory.
1983
1984
1985
                    - 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*):
1986
1987
1988
                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
1989
                weights. In this case though, you should check if using [`~PreTrainedModel.save_pretrained`] and
1990
1991
                [`~PreTrainedModel.from_pretrained`] is not a simpler option.
            cache_dir (`Union[str, os.PathLike]`, *optional*):
1992
1993
                Path to a directory in which a downloaded pretrained model configuration should be cached if the
                standard cache should not be used.
1994
            from_tf (`bool`, *optional*, defaults to `False`):
1995
                Load the model weights from a TensorFlow checkpoint save file (see docstring of
1996
1997
                `pretrained_model_name_or_path` argument).
            from_flax (`bool`, *optional*, defaults to `False`):
1998
                Load the model weights from a Flax checkpoint save file (see docstring of
1999
2000
                `pretrained_model_name_or_path` argument).
            ignore_mismatched_sizes (`bool`, *optional*, defaults to `False`):
2001
2002
2003
                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).
2004
            force_download (`bool`, *optional*, defaults to `False`):
2005
2006
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
2007
            resume_download (`bool`, *optional*, defaults to `False`):
2008
2009
                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
                file exists.
2010
            proxies (`Dict[str, str]`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
2011
2012
                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.
2013
            output_loading_info(`bool`, *optional*, defaults to `False`):
Sylvain Gugger's avatar
Sylvain Gugger committed
2014
                Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
2015
            local_files_only(`bool`, *optional*, defaults to `False`):
Stas Bekman's avatar
Stas Bekman committed
2016
                Whether or not to only look at local files (i.e., do not try to download the model).
2017
            token (`str` or `bool`, *optional*):
2018
2019
                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`).
2020
            revision (`str`, *optional*, defaults to `"main"`):
Julien Chaumond's avatar
Julien Chaumond committed
2021
                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
2022
                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
Julien Chaumond's avatar
Julien Chaumond committed
2023
                identifier allowed by git.
2024
2025
2026
2027
2028
2029
2030

                <Tip>

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

                </Tip>

2031
            mirror (`str`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
2032
2033
2034
                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.
2035
            _fast_init(`bool`, *optional*, defaults to `True`):
2036
2037
                Whether or not to disable fast initialization.

2038
2039
                <Tip warning={true}>

Sylvain Gugger's avatar
Sylvain Gugger committed
2040
2041
2042
                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.
2043

2044
                </Tip>
2045

2046
2047
2048
            > Parameters for big model inference

            low_cpu_mem_usage(`bool`, *optional*):
2049
2050
2051
                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*):
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
                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>

2073
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]` or `int` or `torch.device`, *optional*):
2074
2075
                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
2076
2077
2078
                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.
2079

2080
2081
                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
2082
                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
2083
2084
2085
            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.
2086
2087
            offload_folder (`str` or `os.PathLike`, *optional*):
                If the `device_map` contains any value `"disk"`, the folder where we will offload weights.
2088
            offload_state_dict (`bool`, *optional*):
2089
                If `True`, will temporarily offload the CPU state dict to the hard drive to avoid getting out of CPU
2090
2091
                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.
2092
2093
2094
2095
2096
2097
            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
                install `bitsandbytes` compiled with your CUDA version by running `pip install -i
                https://test.pypi.org/simple/ bitsandbytes-cudaXXX` where XXX is your CUDA version (e.g. 11.6 = 116).
                Make also sure that you have enough GPU RAM to store half of the model size since the 8bit modules are
                not compiled and adapted for CPUs.
2098
2099
2100
            quantization_config (`Dict`, *optional*):
                A dictionary of configuration parameters for the `bitsandbytes` library and loading the model using
                advanced features such as offloading in fp32 on CPU or on disk.
2101
2102
2103
            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.
2104
2105
2106
            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`.
2107
2108
2109
            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`.
2110

2111
            kwargs (remaining dictionary of keyword arguments, *optional*):
2112
                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
2113
                `output_attentions=True`). Behaves differently depending on whether a `config` is provided or
2114
2115
                automatically loaded:

2116
2117
                    - 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
2118
                      already been done)
2119
                    - If a configuration is not provided, `kwargs` will be first passed to the configuration class
Sylvain Gugger's avatar
Sylvain Gugger committed
2120
2121
2122
2123
                      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.
2124
2125
2126

        <Tip>

Sylvain Gugger's avatar
Sylvain Gugger committed
2127
2128
        Activate the special ["offline-mode"](https://huggingface.co/transformers/installation.html#offline-mode) to
        use this method in a firewalled environment.
2129
2130
2131
2132
2133
2134
2135

        </Tip>

        Examples:

        ```python
        >>> from transformers import BertConfig, BertModel
Sylvain Gugger's avatar
Sylvain Gugger committed
2136

2137
        >>> # Download model and configuration from huggingface.co and cache.
Sylvain Gugger's avatar
Sylvain Gugger committed
2138
        >>> model = BertModel.from_pretrained("bert-base-uncased")
2139
        >>> # Model was saved using *save_pretrained('./test/saved_model/')* (for example purposes, not runnable).
Sylvain Gugger's avatar
Sylvain Gugger committed
2140
        >>> model = BertModel.from_pretrained("./test/saved_model/")
2141
        >>> # Update configuration during loading.
Sylvain Gugger's avatar
Sylvain Gugger committed
2142
        >>> model = BertModel.from_pretrained("bert-base-uncased", output_attentions=True)
2143
2144
        >>> 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
2145
2146
        >>> 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)
2147
        >>> # Loading from a Flax checkpoint file instead of a PyTorch model (slower)
Sylvain Gugger's avatar
Sylvain Gugger committed
2148
        >>> model = BertModel.from_pretrained("bert-base-uncased", from_flax=True)
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
        ```

        * `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

        """
2167
2168
        state_dict = kwargs.pop("state_dict", None)
        from_tf = kwargs.pop("from_tf", False)
2169
        from_flax = kwargs.pop("from_flax", False)
2170
2171
2172
        resume_download = kwargs.pop("resume_download", False)
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
2173
        use_auth_token = kwargs.pop("use_auth_token", None)
2174
        trust_remote_code = kwargs.pop("trust_remote_code", None)
Sylvain Gugger's avatar
Sylvain Gugger committed
2175
        _ = kwargs.pop("mirror", None)
2176
2177
        from_pipeline = kwargs.pop("_from_pipeline", None)
        from_auto_class = kwargs.pop("_from_auto", False)
2178
        _fast_init = kwargs.pop("_fast_init", True)
2179
        torch_dtype = kwargs.pop("torch_dtype", None)
2180
2181
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", None)
        device_map = kwargs.pop("device_map", None)
2182
        max_memory = kwargs.pop("max_memory", None)
2183
        offload_folder = kwargs.pop("offload_folder", None)
2184
2185
        offload_state_dict = kwargs.pop("offload_state_dict", False)
        load_in_8bit = kwargs.pop("load_in_8bit", False)
2186
        load_in_4bit = kwargs.pop("load_in_4bit", False)
2187
        quantization_config = kwargs.pop("quantization_config", None)
2188
        subfolder = kwargs.pop("subfolder", "")
2189
        commit_hash = kwargs.pop("_commit_hash", None)
2190
        variant = kwargs.pop("variant", None)
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203

        if use_auth_token is not None:
            warnings.warn(
                "The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers.", FutureWarning
            )
            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 use_safetensors is None and not is_safetensors_available():
            use_safetensors = False
2204

2205
2206
2207
2208
2209
        if is_bitsandbytes_available():
            is_8bit_serializable = version.parse(importlib_metadata.version("bitsandbytes")) > version.parse("0.37.2")
        else:
            is_8bit_serializable = False

2210
2211
2212
2213
2214
        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."
            )
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234

        # 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}

2235
2236
2237
2238
2239
2240
2241
        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:
2242
            if device_map is not None:
2243
2244
                # The max memory utils require PyTorch >= 1.10 to have torch.cuda.mem_get_info.
                require_version_core("torch>=1.10")
2245
2246
2247
2248
2249
2250
2251
2252
2253

            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`"
                )
2254

2255
2256
        if quantization_config is None:
            quantization_config, kwargs = BitsAndBytesConfig.from_dict(
2257
2258
2259
                config_dict={"load_in_8bit": load_in_8bit, "load_in_4bit": load_in_4bit},
                return_unused_kwargs=True,
                **kwargs,
2260
2261
2262
            )
        elif quantization_config is not None:
            load_in_8bit = quantization_config.load_in_8bit
2263
            load_in_4bit = quantization_config.load_in_4bit
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274

            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."
                )

2275
        if load_in_8bit or load_in_4bit:
2276
2277
2278
2279
2280
2281
            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"
                    " pip install bitsandbytes` "
                )
2282
2283

            if torch_dtype is None:
2284
                # We force the `dtype` to be float16, this is a requirement from `bitsandbytes`
2285
                logger.info(
2286
                    f"Overriding torch_dtype={torch_dtype} with `torch_dtype=torch.float16` due to "
2287
                    "requirements of `bitsandbytes` to enable model loading in 8-bit or 4-bit. "
2288
2289
                    "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."
2290
                )
2291
                torch_dtype = torch.float16
2292

2293
            if device_map is None:
2294
2295
2296
2297
2298
2299
2300
2301
                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(
                    "The device_map was not initialized."
                    "Setting device_map to {'':torch.cuda.current_device()}."
                    "If you want to use the model for inference, please set device_map ='auto' "
2302
                )
2303
2304
2305
                if low_cpu_mem_usage is None:
                    low_cpu_mem_usage = True

2306
2307
            if from_tf or from_flax:
                raise ValueError(
2308
                    "Converting into 4-bit or 8-bit weights from tf/flax weights is currently not supported, please make"
2309
2310
2311
                    " sure the weights are in PyTorch format."
                )

2312
        from_pt = not (from_tf | from_flax)
2313
2314
2315
2316

        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
2317

2318
2319
2320
2321
        if is_offline_mode() and not local_files_only:
            logger.info("Offline mode: forcing local_files_only=True")
            local_files_only = True

2322
2323
2324
        # 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
2325
            config, model_kwargs = cls.config_class.from_pretrained(
2326
2327
2328
                config_path,
                cache_dir=cache_dir,
                return_unused_kwargs=True,
2329
                force_download=force_download,
2330
                resume_download=resume_download,
2331
                proxies=proxies,
2332
                local_files_only=local_files_only,
2333
                token=token,
Julien Chaumond's avatar
Julien Chaumond committed
2334
                revision=revision,
2335
                subfolder=subfolder,
2336
2337
                _from_auto=from_auto_class,
                _from_pipeline=from_pipeline,
2338
                **kwargs,
2339
2340
2341
            )
        else:
            model_kwargs = kwargs
2342

2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
        if is_8bit_serializable and quantization_config is not None and load_in_8bit:
            if hasattr(config, "quantization_config"):
                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
        elif is_8bit_serializable and not load_in_8bit and hasattr(config, "quantization_config"):
            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:
2366
2367
                if torch_dtype is None:
                    torch_dtype = torch.float16
2368
                if device_map is None:
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
                    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(
                        "The device_map was not initialized."
                        "Setting device_map to {'':torch.cuda.current_device()}."
                        "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
2380
2381
2382
2383
2384
2385
2386
2387

        elif not is_8bit_serializable and not load_in_8bit and hasattr(config, "quantization_config"):
            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`."
            )

2388
2389
2390
        if commit_hash is None:
            commit_hash = getattr(config, "_commit_hash", None)

Sylvain Gugger's avatar
Sylvain Gugger committed
2391
2392
2393
2394
        # 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
2395
        # Load model
Yih-Dar's avatar
Yih-Dar committed
2396
2397
        loading_info = None

2398
2399
2400
2401
        # Keep in fp32 modules
        keep_in_fp32_modules = None
        use_keep_in_fp32_modules = False

thomwolf's avatar
thomwolf committed
2402
        if pretrained_model_name_or_path is not None:
2403
            pretrained_model_name_or_path = str(pretrained_model_name_or_path)
2404
2405
            is_local = os.path.isdir(pretrained_model_name_or_path)
            if is_local:
2406
2407
2408
                if from_tf and os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, TF_WEIGHTS_NAME + ".index")
                ):
2409
                    # Load from a TF 1.0 checkpoint in priority if from_tf
2410
2411
2412
2413
                    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)
                ):
2414
                    # Load from a TF 2.0 checkpoint in priority if from_tf
2415
2416
2417
2418
                    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)
                ):
2419
                    # Load from a Flax checkpoint in priority if from_flax
2420
                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)
2421
                elif use_safetensors is not False and os.path.isfile(
2422
                    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant))
2423
2424
                ):
                    # Load from a safetensors checkpoint
2425
2426
2427
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant)
                    )
2428
                elif use_safetensors is not False and os.path.isfile(
2429
2430
2431
                    os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
                    )
2432
2433
                ):
                    # Load from a sharded safetensors checkpoint
2434
2435
2436
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
                    )
2437
                    is_sharded = True
2438
2439
2440
                elif os.path.isfile(
                    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_NAME, variant))
                ):
thomwolf's avatar
thomwolf committed
2441
                    # Load from a PyTorch checkpoint
2442
2443
2444
2445
2446
2447
                    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
2448
                    # Load from a sharded PyTorch checkpoint
2449
2450
2451
                    archive_file = os.path.join(
                        pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_INDEX_NAME, variant)
                    )
Sylvain Gugger's avatar
Sylvain Gugger committed
2452
                    is_sharded = True
2453
2454
                # At this stage we don't have a weight file so we will raise an error.
                elif os.path.isfile(
2455
2456
                    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)):
2457
                    raise EnvironmentError(
2458
2459
2460
                        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."
2461
                    )
2462
                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)):
2463
                    raise EnvironmentError(
2464
2465
2466
                        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."
2467
                    )
2468
2469
2470
2471
2472
                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
2473
                else:
2474
                    raise EnvironmentError(
2475
2476
2477
                        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}."
2478
                    )
2479
            elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path)):
2480
                archive_file = pretrained_model_name_or_path
2481
                is_local = True
2482
            elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path + ".index")):
2483
2484
2485
2486
2487
                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."
                    )
2488
                archive_file = os.path.join(subfolder, pretrained_model_name_or_path + ".index")
2489
                is_local = True
2490
            elif is_remote_url(pretrained_model_name_or_path):
2491
                filename = pretrained_model_name_or_path
2492
                resolved_archive_file = download_url(pretrained_model_name_or_path)
2493
            else:
2494
2495
2496
2497
2498
                # set correct filename
                if from_tf:
                    filename = TF2_WEIGHTS_NAME
                elif from_flax:
                    filename = FLAX_WEIGHTS_NAME
2499
                elif use_safetensors is not False:
2500
                    filename = _add_variant(SAFE_WEIGHTS_NAME, variant)
2501
                else:
2502
                    filename = _add_variant(WEIGHTS_NAME, variant)
2503

2504
2505
                try:
                    # Load from URL or cache if already cached
2506
2507
2508
2509
2510
2511
                    cached_file_kwargs = {
                        "cache_dir": cache_dir,
                        "force_download": force_download,
                        "proxies": proxies,
                        "resume_download": resume_download,
                        "local_files_only": local_files_only,
2512
                        "use_auth_token": token,
2513
2514
2515
2516
2517
2518
                        "user_agent": user_agent,
                        "revision": revision,
                        "subfolder": subfolder,
                        "_raise_exceptions_for_missing_entries": False,
                        "_commit_hash": commit_hash,
                    }
2519
                    resolved_archive_file = cached_file(pretrained_model_name_or_path, filename, **cached_file_kwargs)
2520

2521
                    # Since we set _raise_exceptions_for_missing_entries=False, we don't get an exception but a None
2522
                    # result when internet is up, the repo and revision exist, but the file does not.
2523
                    if resolved_archive_file is None and filename == _add_variant(SAFE_WEIGHTS_NAME, variant):
2524
2525
                        # Maybe the checkpoint is sharded, we try to grab the index name in this case.
                        resolved_archive_file = cached_file(
2526
2527
2528
                            pretrained_model_name_or_path,
                            _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant),
                            **cached_file_kwargs,
2529
2530
2531
                        )
                        if resolved_archive_file is not None:
                            is_sharded = True
2532
2533
2534
2535
                        elif use_safetensors:
                            raise EnvironmentError(
                                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`."
                            )
2536
2537
                        else:
                            # This repo has no safetensors file of any kind, we switch to PyTorch.
2538
                            filename = _add_variant(WEIGHTS_NAME, variant)
2539
                            resolved_archive_file = cached_file(
2540
                                pretrained_model_name_or_path, filename, **cached_file_kwargs
2541
                            )
2542
                    if resolved_archive_file is None and filename == _add_variant(WEIGHTS_NAME, variant):
Sylvain Gugger's avatar
Sylvain Gugger committed
2543
                        # Maybe the checkpoint is sharded, we try to grab the index name in this case.
2544
                        resolved_archive_file = cached_file(
2545
2546
2547
                            pretrained_model_name_or_path,
                            _add_variant(WEIGHTS_INDEX_NAME, variant),
                            **cached_file_kwargs,
2548
                        )
2549
2550
2551
                        if resolved_archive_file is not None:
                            is_sharded = True
                    if resolved_archive_file is None:
Sylvain Gugger's avatar
Sylvain Gugger committed
2552
2553
2554
2555
2556
                        # 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,
2557
                            "use_auth_token": token,
Sylvain Gugger's avatar
Sylvain Gugger committed
2558
2559
2560
                        }
                        if has_file(pretrained_model_name_or_path, TF2_WEIGHTS_NAME, **has_file_kwargs):
                            raise EnvironmentError(
Sylvain Gugger's avatar
Sylvain Gugger committed
2561
                                f"{pretrained_model_name_or_path} does not appear to have a file named"
2562
2563
                                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
2564
2565
2566
                            )
                        elif has_file(pretrained_model_name_or_path, FLAX_WEIGHTS_NAME, **has_file_kwargs):
                            raise EnvironmentError(
Sylvain Gugger's avatar
Sylvain Gugger committed
2567
                                f"{pretrained_model_name_or_path} does not appear to have a file named"
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
                                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
2578
2579
2580
                            )
                        else:
                            raise EnvironmentError(
2581
2582
2583
                                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
2584
                            )
2585
2586
2587
2588
2589
2590
                except EnvironmentError:
                    # Raise any environment error raise by `cached_file`. It will have a helpful error message adapted
                    # to the original exception.
                    raise
                except Exception:
                    # For any other exception, we throw a generic error.
2591
                    raise EnvironmentError(
2592
2593
2594
                        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"
2595
2596
                        f" directory containing a file named {_add_variant(WEIGHTS_NAME, variant)},"
                        f" {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME} or {FLAX_WEIGHTS_NAME}."
2597
                    )
2598

2599
            if is_local:
2600
                logger.info(f"loading weights file {archive_file}")
2601
                resolved_archive_file = archive_file
2602
            else:
2603
                logger.info(f"loading weights file {filename} from cache at {resolved_archive_file}")
2604
        else:
thomwolf's avatar
thomwolf committed
2605
            resolved_archive_file = None
2606

Sylvain Gugger's avatar
Sylvain Gugger committed
2607
2608
        # We'll need to download and cache each checkpoint shard if the checkpoint is sharded.
        if is_sharded:
2609
            # 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
2610
2611
2612
2613
2614
2615
2616
2617
            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,
2618
                use_auth_token=token,
Sylvain Gugger's avatar
Sylvain Gugger committed
2619
2620
                user_agent=user_agent,
                revision=revision,
2621
                subfolder=subfolder,
2622
                _commit_hash=commit_hash,
Sylvain Gugger's avatar
Sylvain Gugger committed
2623
2624
            )

2625
2626
        # load pt weights early so that we know which dtype to init the model under
        if from_pt:
2627
            if not is_sharded and state_dict is None:
Sylvain Gugger's avatar
Sylvain Gugger committed
2628
2629
                # Time to load the checkpoint
                state_dict = load_state_dict(resolved_archive_file)
2630

2631
2632
2633
            # 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
2634
            #    weights entry that is of a floating type - we assume all floating dtype weights are of the same dtype
2635
2636
            # we also may have config.torch_dtype available, but we won't rely on it till v5
            dtype_orig = None
2637

2638
2639
2640
            if torch_dtype is not None:
                if isinstance(torch_dtype, str):
                    if torch_dtype == "auto":
2641
2642
2643
                        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
2644
                        else:
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
                            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"
                            )
2657
2658
                    else:
                        raise ValueError(
2659
                            f'`torch_dtype` can be either `torch.dtype` or `"auto"`, but received {torch_dtype}'
2660
2661
2662
                        )
                dtype_orig = cls._set_default_torch_dtype(torch_dtype)

2663
2664
            # Check if `_keep_in_fp32_modules` is not None
            use_keep_in_fp32_modules = (
2665
2666
2667
                (cls._keep_in_fp32_modules is not None)
                and is_accelerate_available()
                and (torch_dtype == torch.float16 or load_in_4bit or load_in_8bit)
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
            )
            if (
                (cls._keep_in_fp32_modules is not None)
                and not is_accelerate_available()
                and torch_dtype == torch.float16
            ):
                logger.warning(
                    "For stability purposes, it is recommended to have accelerate installed when using this model in"
                    " torch.float16, please install it with `pip install accelerate`"
                )

2679
2680
2681
            if is_sharded:
                loaded_state_dict_keys = sharded_metadata["all_checkpoint_keys"]
            else:
2682
                loaded_state_dict_keys = list(state_dict.keys())
2683
            if low_cpu_mem_usage or use_keep_in_fp32_modules:
2684
                state_dict = None
2685

2686
2687
        config.name_or_path = pretrained_model_name_or_path

2688
        # Instantiate model.
2689
2690
        init_contexts = [no_init_weights(_enable=_fast_init)]

2691
2692
2693
2694
        if is_deepspeed_zero3_enabled():
            import deepspeed

            logger.info("Detected DeepSpeed ZeRO-3: activating zero.init() for this model")
2695
            init_contexts = [deepspeed.zero.Init(config_dict_or_path=deepspeed_config())] + init_contexts
2696
        elif load_in_8bit or load_in_4bit or low_cpu_mem_usage:
2697
2698
2699
2700
2701
            init_contexts.append(init_empty_weights())

        with ContextManagers(init_contexts):
            model = cls(config, *model_args, **model_kwargs)

2702
2703
2704
2705
2706
2707
2708
        # Check first if we are `from_pt`
        if use_keep_in_fp32_modules:
            low_cpu_mem_usage = True
            keep_in_fp32_modules = model._keep_in_fp32_modules
        else:
            keep_in_fp32_modules = []

2709
2710
        if load_in_8bit or load_in_4bit:
            from .utils.bitsandbytes import get_keys_to_not_convert, replace_with_bnb_linear
2711

2712
            llm_int8_skip_modules = quantization_config.llm_int8_skip_modules
2713
2714
            load_in_8bit_fp32_cpu_offload = quantization_config.llm_int8_enable_fp32_cpu_offload

2715
2716
            logger.info("Detected 8-bit loading: activating 8-bit loading for this model")

2717
            # We keep some modules such as the lm_head in their original dtype for numerical stability reasons
2718
            if llm_int8_skip_modules is None:
2719
2720
                modules_to_not_convert = get_keys_to_not_convert(model)
            else:
2721
                modules_to_not_convert = llm_int8_skip_modules
2722
2723
2724
2725
2726
2727

            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)

2728
2729
2730
2731
2732
2733
2734
            # 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
2735
                        "`llm_int8_enable_fp32_cpu_offload=True`. Note that these modules will not be "
2736
2737
2738
2739
2740
                        " converted to 8-bit but kept in 32-bit."
                    )

                modules_to_not_convert.extend(keys_on_cpu)

2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
            supports_4bit = version.parse(importlib_metadata.version("bitsandbytes")) >= version.parse("0.39.0")

            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
2751
            )
2752
            # training in 8-bit is only available in 0.37.0+
2753
            model._is_quantized_training_enabled = version.parse(
2754
2755
2756
                importlib_metadata.version("bitsandbytes")
            ) >= version.parse("0.37.0")

2757
2758
2759
            model.config.quantization_config = quantization_config
            model.is_8bit_serializable = is_8bit_serializable

2760
2761
2762
        if load_in_8bit and torch_dtype is None:
            logger.warning(
                "You are loading your model in 8bit but you did not specify a `torch_dtype` attribute."
2763
2764
                "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."
2765
2766
            )

2767
        if isinstance(device_map, str):
2768
            special_dtypes = {}
2769
            if load_in_8bit or load_in_4bit:
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
                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)
                }
            )

2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
            target_dtype = torch_dtype

            if load_in_4bit:
                if version.parse(importlib_metadata.version("accelerate")) > version.parse("0.19.0"):
                    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"
                        " the appropriate device map, you should upgrade your `accelerate` library,"
2797
                        "`pip install --upgrade accelerate` or install it from source to support fp4 auto device map"
2798
2799
2800
2801
2802
                        "calculation. You may encounter unexpected behavior, or pass your own device map"
                    )
            elif load_in_8bit:
                target_dtype = torch.int8

2803
            if model._no_split_modules is None:
2804
2805
2806
2807
                raise ValueError(
                    f"{model.__class__.__name__} does not support `device_map='{device_map}'`. To implement support, the model"
                    "class needs to implement the `_no_split_modules` attribute."
                )
2808
            no_split_modules = model._no_split_modules
2809
2810
2811
2812
2813
            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'."
                )
2814

2815
            kwargs = {"no_split_module_classes": no_split_modules}
2816
2817
2818
2819
2820
2821
2822
            if "special_dtypes" in inspect.signature(infer_auto_device_map).parameters:
                kwargs["special_dtypes"] = special_dtypes
            elif len(special_dtypes) > 0:
                logger.warn(
                    "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`)."
                )
2823
            if device_map != "sequential":
2824
2825
                max_memory = get_balanced_memory(
                    model,
2826
                    dtype=target_dtype,
2827
                    low_zero=(device_map == "balanced_low_0"),
2828
                    max_memory=max_memory,
2829
                    **kwargs,
2830
                )
2831
            kwargs["max_memory"] = max_memory
2832
2833
            # Make sure tied weights are tied before creating the device map.
            model.tie_weights()
2834
            device_map = infer_auto_device_map(model, dtype=target_dtype, **kwargs)
2835

2836
            if load_in_8bit or load_in_4bit:
2837
                # The LM head / tied weights or any last module can stay on disk / CPU
2838
                device_map_without_lm_head = {
2839
                    key: device_map[key] for key in device_map.keys() if key not in modules_to_not_convert
2840
2841
                }
                if "cpu" in device_map_without_lm_head.values() or "disk" in device_map_without_lm_head.values():
2842
2843
2844
                    raise ValueError(
                        """
                        Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit
2845
2846
2847
2848
2849
                        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.
2850
2851
                        """
                    )
2852
2853
                del device_map_without_lm_head

2854
2855
2856
2857
        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
2858
            check_tied_parameters_on_same_device(tied_params, device_map)
2859

2860
        if from_tf:
2861
            if resolved_archive_file.endswith(".index"):
2862
2863
2864
2865
2866
                # 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:
2867
                    from .modeling_tf_pytorch_utils import load_tf2_checkpoint_in_pytorch_model
2868

Yih-Dar's avatar
Yih-Dar committed
2869
2870
2871
                    model, loading_info = load_tf2_checkpoint_in_pytorch_model(
                        model, resolved_archive_file, allow_missing_keys=True, output_loading_info=True
                    )
2872
                except ImportError:
2873
                    logger.error(
Sylvain Gugger's avatar
Sylvain Gugger committed
2874
2875
2876
                        "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."
2877
                    )
2878
                    raise
2879
2880
2881
2882
2883
2884
2885
        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
2886
2887
2888
                    "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."
2889
2890
                )
                raise
2891
        elif from_pt:
2892
2893
2894
2895
            # restore default dtype
            if dtype_orig is not None:
                torch.set_default_dtype(dtype_orig)

Sylvain Gugger's avatar
Sylvain Gugger committed
2896
2897
2898
2899
2900
2901
2902
2903
            (
                model,
                missing_keys,
                unexpected_keys,
                mismatched_keys,
                offload_index,
                error_msgs,
            ) = cls._load_pretrained_model(
2904
2905
2906
2907
2908
2909
2910
2911
2912
                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,
2913
2914
2915
2916
                device_map=device_map,
                offload_folder=offload_folder,
                offload_state_dict=offload_state_dict,
                dtype=torch_dtype,
2917
                is_quantized=(load_in_8bit or load_in_4bit),
2918
                keep_in_fp32_modules=keep_in_fp32_modules,
2919
            )
2920

2921
        model.is_loaded_in_4bit = load_in_4bit
Younes Belkada's avatar
Younes Belkada committed
2922
        model.is_loaded_in_8bit = load_in_8bit
2923
        model.is_quantized = load_in_8bit or load_in_4bit
2924

2925
2926
        # make sure token embedding weights are still tied if needed
        model.tie_weights()
2927

2928
        # Set model in evaluation mode to deactivate DropOut modules by default
2929
2930
        model.eval()

2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
        # If it is a model with generation capabilities, attempt to load the generation config
        if model.can_generate():
            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,
2941
                    token=token,
2942
2943
2944
2945
2946
2947
                    revision=revision,
                    subfolder=subfolder,
                    _from_auto=from_auto_class,
                    _from_pipeline=from_pipeline,
                    **kwargs,
                )
2948
            except (OSError, TypeError):
2949
2950
2951
2952
2953
                logger.info(
                    "Generation config file not found, using a generation config created from the model config."
                )
                pass

2954
2955
        # Dispatch model with hooks on all devices if necessary
        if device_map is not None:
2956
2957
2958
2959
            kwargs = {"device_map": device_map, "offload_dir": offload_folder, "offload_index": offload_index}
            if "skip_keys" in inspect.signature(dispatch_model).parameters:
                kwargs["skip_keys"] = model._skip_keys_device_placement
            dispatch_model(model, **kwargs)
2960

thomwolf's avatar
thomwolf committed
2961
        if output_loading_info:
Yih-Dar's avatar
Yih-Dar committed
2962
2963
2964
2965
2966
2967
2968
            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
2969
2970
            return model, loading_info

2971
2972
        return model

2973
    @classmethod
Sylvain Gugger's avatar
Sylvain Gugger committed
2974
2975
2976
2977
    def _load_pretrained_model(
        cls,
        model,
        state_dict,
2978
        loaded_keys,
Sylvain Gugger's avatar
Sylvain Gugger committed
2979
2980
2981
2982
2983
        resolved_archive_file,
        pretrained_model_name_or_path,
        ignore_mismatched_sizes=False,
        sharded_metadata=None,
        _fast_init=True,
2984
        low_cpu_mem_usage=False,
2985
2986
        device_map=None,
        offload_folder=None,
2987
        offload_state_dict=None,
2988
        dtype=None,
2989
        is_quantized=False,
2990
        keep_in_fp32_modules=None,
2991
    ):
Sylvain Gugger's avatar
Sylvain Gugger committed
2992
        is_safetensors = False
2993
2994
        if is_quantized:
            from .utils.bitsandbytes import set_module_quantized_tensor_to_device
2995

Sylvain Gugger's avatar
Sylvain Gugger committed
2996
        if device_map is not None and "disk" in device_map.values():
Sylvain Gugger's avatar
Sylvain Gugger committed
2997
2998
2999
3000
3001
            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
3002
3003
                raise ValueError(
                    "The current `device_map` had weights offloaded to the disk. Please provide an `offload_folder`"
Sylvain Gugger's avatar
Sylvain Gugger committed
3004
3005
                    " 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
3006
                )
Sylvain Gugger's avatar
Sylvain Gugger committed
3007
3008
            if offload_folder is not None:
                os.makedirs(offload_folder, exist_ok=True)
3009
3010
3011
            if offload_state_dict is None:
                offload_state_dict = True

3012
        is_sharded_safetensors = is_safetensors and sharded_metadata is not None
3013
        # Retrieve missing & unexpected_keys
3014
3015
        model_state_dict = model.state_dict()
        expected_keys = list(model_state_dict.keys())
3016
3017
        prefix = model.base_model_prefix

Sylvain Gugger's avatar
Sylvain Gugger committed
3018
3019
3020
3021
3022
3023
3024
        def _fix_key(key):
            if "beta" in key:
                return key.replace("beta", "bias")
            if "gamma" in key:
                return key.replace("gamma", "weight")
            return key

3025
        original_loaded_keys = loaded_keys
Sylvain Gugger's avatar
Sylvain Gugger committed
3026
3027
        loaded_keys = [_fix_key(key) for key in loaded_keys]

3028
3029
3030
3031
3032
3033
        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
3034
3035
3036

        # key re-naming operations are never done on the keys
        # that are loaded, but always on the keys of the newly initialized model
3037
3038
        remove_prefix_from_model = not has_prefix_module and expects_prefix_module
        add_prefix_to_model = has_prefix_module and not expects_prefix_module
3039

3040
        if remove_prefix_from_model:
3041
3042
3043
            _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]
3044
        elif add_prefix_to_model:
3045
3046
3047
            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
3048
3049
3050
3051
3052
3053
3054
3055
3056
        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)
3057

Sylvain Gugger's avatar
Sylvain Gugger committed
3058
3059
3060
3061
3062
3063
3064
3065
        model.tie_weights()
        ptrs = collections.defaultdict(list)
        for name, tensor in model.state_dict().items():
            id_tensor = id_tensor_storage(tensor) if tensor.device != torch.device("meta") else id(tensor)
            ptrs[id_tensor].append(name)

        # These are all the pointers of shared tensors.
        tied_params = [names for _, names in ptrs.items() if len(names) > 1]
Sylvain Gugger's avatar
Sylvain Gugger committed
3066
3067

        for group in tied_params:
Sylvain Gugger's avatar
Sylvain Gugger committed
3068
3069
3070
3071
            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
3072
3073
3074
            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]
3075

3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
        # 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]

3086
3087
3088
3089
        # 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
3090
3091
                if key in list(model_state_dict.keys()):
                    key = key
3092
3093
                elif f"{prefix}.{key}" in list(model_state_dict.keys()):
                    key = f"{prefix}.{key}"
Susnato Dhar's avatar
Susnato Dhar committed
3094
                elif key.startswith(prefix) and ".".join(key.split(".")[1:]) in list(model_state_dict.keys()):
3095
3096
                    key = ".".join(key.split(".")[1:])
                param = model_state_dict[key]
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106

                # upcast in fp32 if any
                target_dtype = dtype
                if (
                    keep_in_fp32_modules is not None
                    and dtype == torch.float16
                    and any(module_to_keep_in_fp32 in key for module_to_keep_in_fp32 in keep_in_fp32_modules)
                ):
                    target_dtype = torch.float32

3107
                if param.device == torch.device("meta"):
3108
                    if not (is_quantized):
3109
                        set_module_tensor_to_device(model, key, "cpu", torch.empty(*param.size(), dtype=target_dtype))
3110
                    else:
3111
                        set_module_quantized_tensor_to_device(
3112
3113
                            model, key, "cpu", torch.empty(*param.size(), dtype=target_dtype)
                        )
3114
3115

        # retrieve unintialized modules and initialize before maybe overriding that with the pretrained weights.
3116
        if _fast_init:
3117
3118
3119
3120
3121
3122
3123
3124
3125
            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)
            # This will only initialize submodules that are not marked as initialized by the line above.
            model.apply(model._initialize_weights)
3126

3127
3128
3129
3130
3131
3132
        # Set some modules to fp32 if any
        if keep_in_fp32_modules is not None:
            for name, param in model.named_parameters():
                if any(module_to_keep_in_fp32 in name for module_to_keep_in_fp32 in keep_in_fp32_modules):
                    param = param.to(torch.float32)

3133
3134
3135
        # Make sure we are able to load base models as well as derived models (with heads)
        start_prefix = ""
        model_to_load = model
3136
        if len(cls.base_model_prefix) > 0 and not hasattr(model, cls.base_model_prefix) and has_prefix_module:
3137
            start_prefix = cls.base_model_prefix + "."
3138
        if len(cls.base_model_prefix) > 0 and hasattr(model, cls.base_model_prefix) and not has_prefix_module:
3139
            model_to_load = getattr(model, cls.base_model_prefix)
Sylvain Gugger's avatar
Sylvain Gugger committed
3140
3141
            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):
3142
                raise ValueError(
3143
                    "The state dictionary of the model you are trying to load is corrupted. Are you sure it was "
3144
3145
                    "properly saved?"
                )
3146
3147
            if device_map is not None:
                device_map = {k.replace(f"{cls.base_model_prefix}.", ""): v for k, v in device_map.items()}
3148

3149
3150
3151
3152
3153
3154
3155
3156
        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
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
            mismatched_keys = []
            if ignore_mismatched_sizes:
                for checkpoint_key in loaded_keys:
                    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]
3176
3177
            return mismatched_keys

3178
3179
3180
3181
        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
3182
        if device_map is not None and is_safetensors:
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
            param_device_map = expand_device_map(device_map, original_loaded_keys)

            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
3195
            offload_index = {
3196
3197
                p: {"safetensors_file": f, "weight_name": p, "dtype": str_dtype}
                for p, f in weight_map.items()
Sylvain Gugger's avatar
Sylvain Gugger committed
3198
3199
3200
                if param_device_map[p] == "disk"
            }

3201
3202
3203
3204
3205
        if state_dict is not None:
            # Whole checkpoint
            mismatched_keys = _find_mismatched_keys(
                state_dict,
                model_state_dict,
3206
                original_loaded_keys,
3207
3208
3209
3210
                add_prefix_to_model,
                remove_prefix_from_model,
                ignore_mismatched_sizes,
            )
Sylvain Gugger's avatar
Sylvain Gugger committed
3211
            error_msgs = _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
Sylvain Gugger's avatar
Sylvain Gugger committed
3212
            offload_index = None
Sylvain Gugger's avatar
Sylvain Gugger committed
3213
        else:
3214
3215
            # Sharded checkpoint or whole but low_cpu_mem_usage==True

Sylvain Gugger's avatar
Sylvain Gugger committed
3216
3217
3218
3219
3220
            # 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 = []
3221
            mismatched_keys = []
Sylvain Gugger's avatar
Sylvain Gugger committed
3222
3223
            if not is_safetensors:
                offload_index = {} if device_map is not None and "disk" in device_map.values() else None
3224
3225
3226
3227
3228
3229
3230
            if offload_state_dict:
                state_dict_folder = tempfile.mkdtemp()
                state_dict_index = {}
            else:
                state_dict_folder = None
                state_dict_index = None

3231
            if is_sharded_safetensors:
Sylvain Gugger's avatar
Sylvain Gugger committed
3232
3233
3234
3235
3236
                disk_only_shard_files = get_disk_only_shard_files(device_map, sharded_metadata=sharded_metadata)
                disk_only_shard_files = [os.path.join(folder, f) for f in disk_only_shard_files]
            else:
                disk_only_shard_files = []

3237
3238
            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
3239
            for shard_file in resolved_archive_file:
Sylvain Gugger's avatar
Sylvain Gugger committed
3240
3241
3242
                # 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
3243
                state_dict = load_state_dict(shard_file)
3244

Sylvain Gugger's avatar
Sylvain Gugger committed
3245
3246
                # Mistmatched keys contains tuples key/shape1/shape2 of weights in the checkpoint that have a shape not
                # matching the weights in the model.
3247
3248
3249
                mismatched_keys += _find_mismatched_keys(
                    state_dict,
                    model_state_dict,
3250
                    original_loaded_keys,
3251
3252
3253
3254
                    add_prefix_to_model,
                    remove_prefix_from_model,
                    ignore_mismatched_sizes,
                )
3255
3256

                if low_cpu_mem_usage:
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
                    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,
3269
                        is_quantized=is_quantized,
Sylvain Gugger's avatar
Sylvain Gugger committed
3270
                        is_safetensors=is_safetensors,
3271
                        keep_in_fp32_modules=keep_in_fp32_modules,
3272
                    )
3273
                    error_msgs += new_error_msgs
3274
3275
                else:
                    error_msgs += _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
3276

3277
3278
3279
3280
                # force memory release
                del state_dict
                gc.collect()

3281
            if offload_index is not None and len(offload_index) > 0:
Sylvain Gugger's avatar
Sylvain Gugger committed
3282
3283
3284
                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
3285
3286
3287
3288
3289
3290
                    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
3291
                    offload_index = {f"{prefix}.{key}": value for key, value in offload_index.items()}
Sylvain Gugger's avatar
Sylvain Gugger committed
3292
3293
3294
                if not is_safetensors:
                    save_offload_index(offload_index, offload_folder)
                    offload_index = None
3295
3296
3297

            if offload_state_dict:
                # Load back temporarily offloaded state dict
3298
                load_offloaded_weights(model_to_load, state_dict_index, state_dict_folder)
3299
3300
                shutil.rmtree(state_dict_folder)

3301
3302
        if len(error_msgs) > 0:
            error_msg = "\n\t".join(error_msgs)
3303
3304
3305
3306
            if "size mismatch" in error_msg:
                error_msg += (
                    "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
                )
3307
3308
            raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")

3309
        if is_quantized:
3310
3311
3312
            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]

3313
        if len(unexpected_keys) > 0:
Sylvain Gugger's avatar
Sylvain Gugger committed
3314
3315
3316
            archs = [] if model.config.architectures is None else model.config.architectures
            warner = logger.warn if model.__class__.__name__ in archs else logger.info
            warner(
Sylvain Gugger's avatar
Sylvain Gugger committed
3317
3318
3319
3320
3321
3322
3323
                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)."
3324
3325
3326
3327
3328
            )
        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
3329
3330
3331
                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."
3332
            )
3333
        elif len(mismatched_keys) == 0:
3334
            logger.info(
Sylvain Gugger's avatar
Sylvain Gugger committed
3335
3336
3337
3338
                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."
3339
            )
3340
3341
3342
3343
3344
3345
3346
3347
        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
3348
3349
3350
3351
                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."
3352
            )
3353

Sylvain Gugger's avatar
Sylvain Gugger committed
3354
        return model, missing_keys, unexpected_keys, mismatched_keys, offload_index, error_msgs
3355
3356

    def retrieve_modules_from_names(self, names, add_prefix=False, remove_prefix=False):
3357
        module_keys = {".".join(key.split(".")[:-1]) for key in names}
3358

Patrick von Platen's avatar
Patrick von Platen committed
3359
3360
        # torch.nn.ParameterList is a special case where two parameter keywords
        # are appended to the module name, *e.g.* bert.special_embeddings.0
3361
        module_keys = module_keys.union(
3362
            {".".join(key.split(".")[:-2]) for key in names if len(key) > 0 and key[-1].isdigit()}
3363
        )
Patrick von Platen's avatar
Patrick von Platen committed
3364

3365
3366
3367
3368
        retrieved_modules = []
        # retrieve all modules that has at least one missing weight name
        for name, module in self.named_modules():
            if remove_prefix:
3369
3370
                _prefix = f"{self.base_model_prefix}."
                name = name[len(_prefix) :] if name.startswith(_prefix) else name
3371
            elif add_prefix:
Patrick von Platen's avatar
Patrick von Platen committed
3372
                name = ".".join([self.base_model_prefix, name]) if len(name) > 0 else self.base_model_prefix
3373
3374
3375
3376
3377
3378

            if name in module_keys:
                retrieved_modules.append(module)

        return retrieved_modules

3379
    @staticmethod
3380
    def _load_pretrained_model_low_mem(model, loaded_state_dict_keys, resolved_archive_file, start_prefix=""):
3381
3382
3383
        """
        This is an experimental function that loads the model using ~1.x model size CPU memory

3384
        Before you call it do:
3385

3386
        1. save which state_dict keys are available
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
        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.
        """

3398
3399
3400
3401
        _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
3402

3403
3404
3405
3406
3407
3408
    @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.

3409
3410
3411
3412
3413
3414
        <Tip warning={true}>

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

        </Tip>

3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
        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

3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
    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)

thomwolf's avatar
thomwolf committed
3479

3480
PreTrainedModel.push_to_hub = copy_func(PreTrainedModel.push_to_hub)
3481
3482
3483
3484
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"
    )
3485
3486


thomwolf's avatar
thomwolf committed
3487
class PoolerStartLogits(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
3488
3489
    """
    Compute SQuAD start logits from sequence hidden states.
3490

Sylvain Gugger's avatar
Sylvain Gugger committed
3491
    Args:
3492
3493
        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
3494
3495
3496
    """

    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
3497
        super().__init__()
thomwolf's avatar
thomwolf committed
3498
3499
        self.dense = nn.Linear(config.hidden_size, 1)

Sylvain Gugger's avatar
Sylvain Gugger committed
3500
3501
3502
3503
3504
    def forward(
        self, hidden_states: torch.FloatTensor, p_mask: Optional[torch.FloatTensor] = None
    ) -> torch.FloatTensor:
        """
        Args:
3505
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
3506
                The final hidden states of the model.
3507
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
3508
3509
                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
3510
3511

        Returns:
3512
            `torch.FloatTensor`: The start logits for SQuAD.
thomwolf's avatar
thomwolf committed
3513
        """
thomwolf's avatar
thomwolf committed
3514
3515
3516
        x = self.dense(hidden_states).squeeze(-1)

        if p_mask is not None:
Lysandre Debut's avatar
Lysandre Debut committed
3517
            if get_parameter_dtype(self) == torch.float16:
3518
3519
3520
                x = x * (1 - p_mask) - 65500 * p_mask
            else:
                x = x * (1 - p_mask) - 1e30 * p_mask
thomwolf's avatar
thomwolf committed
3521
3522
3523
3524
3525
3526

        return x


class PoolerEndLogits(nn.Module):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
3527
    Compute SQuAD end logits from sequence hidden states.
3528

Sylvain Gugger's avatar
Sylvain Gugger committed
3529
    Args:
3530
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
3531
3532
            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
3533
3534
3535
    """

    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
3536
        super().__init__()
thomwolf's avatar
thomwolf committed
3537
3538
3539
3540
3541
        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
3542
3543
3544
3545
3546
3547
3548
3549
3550
    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:
3551
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
3552
                The final hidden states of the model.
3553
            start_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
3554
                The hidden states of the first tokens for the labeled span.
3555
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
3556
                The position of the first token for the labeled span.
3557
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
3558
3559
                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
3560

3561
        <Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
3562

Stas Bekman's avatar
Stas Bekman committed
3563
3564
        One of `start_states` or `start_positions` should be not `None`. If both are set, `start_positions` overrides
        `start_states`.
3565
3566

        </Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
3567
3568

        Returns:
3569
            `torch.FloatTensor`: The end logits for SQuAD.
thomwolf's avatar
thomwolf committed
3570
        """
3571
3572
3573
        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
3574
        if start_positions is not None:
3575
            slen, hsz = hidden_states.shape[-2:]
3576
3577
3578
            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
3579
3580
3581
3582
3583
3584
3585

        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
3586
            if get_parameter_dtype(self) == torch.float16:
3587
3588
3589
                x = x * (1 - p_mask) - 65500 * p_mask
            else:
                x = x * (1 - p_mask) - 1e30 * p_mask
thomwolf's avatar
thomwolf committed
3590
3591
3592
3593
3594

        return x


class PoolerAnswerClass(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
3595
3596
3597
3598
    """
    Compute SQuAD 2.0 answer class from classification and start tokens hidden states.

    Args:
3599
3600
        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
3601
    """
3602

thomwolf's avatar
thomwolf committed
3603
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
3604
        super().__init__()
thomwolf's avatar
thomwolf committed
3605
3606
3607
3608
        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
3609
3610
3611
3612
3613
3614
3615
    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:
3616
3617
        """
        Args:
3618
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Sylvain Gugger's avatar
Sylvain Gugger committed
3619
                The final hidden states of the model.
3620
            start_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
3621
                The hidden states of the first tokens for the labeled span.
3622
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
3623
                The position of the first token for the labeled span.
3624
3625
3626
3627
            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
3628

Stas Bekman's avatar
Stas Bekman committed
3629
3630
        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
3631

3632
        </Tip>
Sylvain Gugger's avatar
Sylvain Gugger committed
3633
3634

        Returns:
3635
            `torch.FloatTensor`: The SQuAD 2.0 answer class.
thomwolf's avatar
thomwolf committed
3636
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
3637
        # No dependency on end_feature so that we can obtain one single `cls_logits` for each sample.
3638
        hsz = hidden_states.shape[-1]
3639
3640
3641
        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
3642
        if start_positions is not None:
3643
3644
            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
3645
3646

        if cls_index is not None:
3647
3648
            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
3649
        else:
3650
            cls_token_state = hidden_states[:, -1, :]  # shape (bsz, hsz)
thomwolf's avatar
thomwolf committed
3651
3652
3653
3654
3655
3656
3657
3658

        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


3659
3660
3661
@dataclass
class SquadHeadOutput(ModelOutput):
    """
3662
    Base class for outputs of question answering models using a [`~modeling_utils.SQuADHead`].
3663
3664

    Args:
3665
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned if both `start_positions` and `end_positions` are provided):
Sylvain Gugger's avatar
Sylvain Gugger committed
3666
3667
            Classification loss as the sum of start token, end token (and is_impossible if provided) classification
            losses.
3668
        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):
3669
            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
3670
        start_top_index (`torch.LongTensor` of shape `(batch_size, config.start_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
3671
            Indices for the top config.start_n_top start token possibilities (beam-search).
3672
3673
        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
3674
            (beam-search).
3675
3676
3677
3678
        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.
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689

    """

    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
3690
class SQuADHead(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
3691
3692
    r"""
    A SQuAD head inspired by XLNet.
3693

Sylvain Gugger's avatar
Sylvain Gugger committed
3694
    Args:
3695
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
3696
3697
            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
3698
    """
3699

thomwolf's avatar
thomwolf committed
3700
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
3701
        super().__init__()
thomwolf's avatar
thomwolf committed
3702
3703
3704
3705
3706
3707
3708
        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
3709
    @replace_return_docstrings(output_type=SquadHeadOutput, config_class=PretrainedConfig)
3710
    def forward(
3711
        self,
Sylvain Gugger's avatar
Sylvain Gugger committed
3712
3713
3714
3715
3716
3717
        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,
3718
        return_dict: bool = False,
Sylvain Gugger's avatar
Sylvain Gugger committed
3719
3720
    ) -> Union[SquadHeadOutput, Tuple[torch.FloatTensor]]:
        """
Lysandre's avatar
Lysandre committed
3721
        Args:
3722
            hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Lysandre's avatar
Lysandre committed
3723
                Final hidden states of the model on the sequence tokens.
3724
            start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Lysandre's avatar
Lysandre committed
3725
                Positions of the first token for the labeled span.
3726
            end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Lysandre's avatar
Lysandre committed
3727
                Positions of the last token for the labeled span.
3728
3729
3730
            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
3731
                Whether the question has a possible answer in the paragraph or not.
3732
            p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
3733
3734
                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
                should be masked.
3735
            return_dict (`bool`, *optional*, defaults to `False`):
3736
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
Sylvain Gugger's avatar
Sylvain Gugger committed
3737

Lysandre's avatar
Lysandre committed
3738
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
3739
        """
thomwolf's avatar
thomwolf committed
3740
        start_logits = self.start_logits(hidden_states, p_mask=p_mask)
thomwolf's avatar
thomwolf committed
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763

        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
3764

3765
            return SquadHeadOutput(loss=total_loss) if return_dict else (total_loss,)
thomwolf's avatar
thomwolf committed
3766
3767
3768
3769

        else:
            # during inference, compute the end logits based on beam search
            bsz, slen, hsz = hidden_states.size()
3770
            start_log_probs = nn.functional.softmax(start_logits, dim=-1)  # shape (bsz, slen)
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781

            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
3782
3783
            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)
3784
            end_log_probs = nn.functional.softmax(end_logits, dim=1)  # shape (bsz, slen, start_n_top)
thomwolf's avatar
thomwolf committed
3785

3786
3787
3788
            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
3789
3790
3791
3792
3793
3794
            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)

3795
            if not return_dict:
3796
3797
3798
3799
3800
3801
3802
3803
3804
                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
3805
3806
3807


class SequenceSummary(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
3808
3809
3810
3811
    r"""
    Compute a single vector summary of a sequence hidden states.

    Args:
3812
        config ([`PretrainedConfig`]):
Sylvain Gugger's avatar
Sylvain Gugger committed
3813
3814
            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
3815

3816
            - **summary_type** (`str`) -- The method to use to make this summary. Accepted values are:
Sylvain Gugger's avatar
Sylvain Gugger committed
3817

3818
3819
3820
3821
3822
                - `"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
3823

3824
            - **summary_use_proj** (`bool`) -- Add a projection after the vector extraction.
Sylvain Gugger's avatar
Sylvain Gugger committed
3825
3826
3827
3828
3829
3830
            - **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
3831
    """
3832

3833
    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
3834
        super().__init__()
thomwolf's avatar
thomwolf committed
3835

3836
        self.summary_type = getattr(config, "summary_type", "last")
3837
        if self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
3838
3839
3840
3841
3842
            # 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
3843
        self.summary = Identity()
3844
3845
        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:
3846
                num_classes = config.num_labels
thomwolf's avatar
thomwolf committed
3847
3848
3849
3850
            else:
                num_classes = config.hidden_size
            self.summary = nn.Linear(config.hidden_size, num_classes)

3851
        activation_string = getattr(config, "summary_activation", None)
Lysandre's avatar
Lysandre committed
3852
        self.activation: Callable = get_activation(activation_string) if activation_string else Identity()
thomwolf's avatar
thomwolf committed
3853

thomwolf's avatar
thomwolf committed
3854
        self.first_dropout = Identity()
3855
        if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
3856
3857
            self.first_dropout = nn.Dropout(config.summary_first_dropout)

thomwolf's avatar
thomwolf committed
3858
        self.last_dropout = Identity()
3859
        if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
3860
            self.last_dropout = nn.Dropout(config.summary_last_dropout)
thomwolf's avatar
thomwolf committed
3861

Sylvain Gugger's avatar
Sylvain Gugger committed
3862
3863
3864
3865
3866
3867
3868
    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:
3869
            hidden_states (`torch.FloatTensor` of shape `[batch_size, seq_len, hidden_size]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
3870
                The hidden states of the last layer.
3871
            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
3872
                Used if `summary_type == "cls_index"` and takes the last token of the sequence as classification token.
Sylvain Gugger's avatar
Sylvain Gugger committed
3873
3874

        Returns:
3875
            `torch.FloatTensor`: The summary of the sequence hidden states.
thomwolf's avatar
thomwolf committed
3876
        """
3877
        if self.summary_type == "last":
thomwolf's avatar
thomwolf committed
3878
            output = hidden_states[:, -1]
3879
        elif self.summary_type == "first":
thomwolf's avatar
thomwolf committed
3880
            output = hidden_states[:, 0]
3881
        elif self.summary_type == "mean":
thomwolf's avatar
thomwolf committed
3882
            output = hidden_states.mean(dim=1)
3883
        elif self.summary_type == "cls_index":
thomwolf's avatar
thomwolf committed
3884
            if cls_index is None:
Lysandre's avatar
Lysandre committed
3885
3886
3887
3888
3889
                cls_index = torch.full_like(
                    hidden_states[..., :1, :],
                    hidden_states.shape[-2] - 1,
                    dtype=torch.long,
                )
thomwolf's avatar
thomwolf committed
3890
            else:
thomwolf's avatar
thomwolf committed
3891
                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
3892
                cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
thomwolf's avatar
thomwolf committed
3893
            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
3894
3895
            output = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, XX, hidden_size)
        elif self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
3896
3897
            raise NotImplementedError

3898
        output = self.first_dropout(output)
thomwolf's avatar
thomwolf committed
3899
3900
        output = self.summary(output)
        output = self.activation(output)
3901
        output = self.last_dropout(output)
thomwolf's avatar
thomwolf committed
3902
3903
3904
3905

        return output


3906
def unwrap_model(model: nn.Module) -> nn.Module:
3907
3908
3909
3910
    """
    Recursively unwraps a model from potential containers (as used in distributed training).

    Args:
3911
        model (`torch.nn.Module`): The model to unwrap.
3912
3913
3914
3915
3916
3917
    """
    # 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
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940


def expand_device_map(device_map, param_names):
    """
    Expand a device map to return the correspondance parameter name to device.
    """
    new_device_map = {}
    for module, device in device_map.items():
        new_device_map.update({p: device for p in param_names if p == module or p.startswith(f"{module}.")})
    return new_device_map


def get_disk_only_shard_files(device_map, sharded_metadata):
    """
    Returns the list of shard files containing only weights offloaded to disk.
    """
    files_content = collections.defaultdict(list)
    for weight_name, filename in sharded_metadata["weight_map"].items():
        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"}]