import collections
import gc
import inspect
import json
import os
import re
import shutil
import tempfile
import warnings
from contextlib import contextmanager
from dataclasses import dataclass
from functools import partial
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import torch
from packaging import version
from torch import Tensor, nn
from torch.nn import CrossEntropyLoss
from .activations import get_activation
from .configuration_utils import PretrainedConfig
from .deepspeed import deepspeed_config, is_deepspeed_zero3_enabled
from .dynamic_module_utils import custom_object_save
from .generation import GenerationConfig, GenerationMixin
from .pytorch_utils import (
Conv1D,
apply_chunking_to_forward,
find_pruneable_heads_and_indices,
id_tensor_storage,
prune_conv1d_layer,
prune_layer,
prune_linear_layer,
)
from .utils import (
DUMMY_INPUTS,
FLAX_WEIGHTS_NAME,
SAFE_WEIGHTS_INDEX_NAME,
SAFE_WEIGHTS_NAME,
TF2_WEIGHTS_NAME,
TF_WEIGHTS_NAME,
WEIGHTS_INDEX_NAME,
WEIGHTS_NAME,
ContextManagers,
ModelOutput,
PushToHubMixin,
cached_file,
copy_func,
download_url,
has_file,
is_accelerate_available,
is_bitsandbytes_available,
is_offline_mode,
is_optimum_available,
is_remote_url,
is_safetensors_available,
is_torch_tpu_available,
logging,
replace_return_docstrings,
)
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
from .utils.quantization_config import BitsAndBytesConfig
from .utils.versions import require_version_core
XLA_USE_BF16 = os.environ.get("XLA_USE_BF16", "0").upper()
XLA_DOWNCAST_BF16 = os.environ.get("XLA_DOWNCAST_BF16", "0").upper()
if is_accelerate_available():
from accelerate import __version__ as accelerate_version
from accelerate import dispatch_model, infer_auto_device_map, init_empty_weights
from accelerate.utils import (
find_tied_parameters,
load_offloaded_weights,
offload_weight,
save_offload_index,
set_module_tensor_to_device,
)
if version.parse(accelerate_version) > version.parse("0.11.0"):
from accelerate.utils import get_balanced_memory
else:
get_balanced_memory = None
if version.parse(accelerate_version) > version.parse("0.19.0"):
from accelerate.utils import check_tied_parameters_on_same_device
else:
check_tied_parameters_on_same_device = None
else:
find_tied_parameters = None
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
logger = logging.get_logger(__name__)
_init_weights = True
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
@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
old_init_weights = _init_weights
if _enable:
_init_weights = False
try:
yield
finally:
_init_weights = old_init_weights
try:
from torch.nn import Identity
except ImportError:
class Identity(nn.Module):
r"""A placeholder identity operator that is argument-insensitive."""
def __init__(self, *args, **kwargs):
super().__init__()
def forward(self, input):
return input
def get_parameter_device(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
try:
return next(parameter.parameters()).device
except StopIteration:
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
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.
"""
try:
return next(parameter.parameters()).dtype
except StopIteration:
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
def get_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
"""
Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.
"""
last_dtype = None
for t in parameter.parameters():
last_dtype = t.dtype
if t.is_floating_point():
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:
return torch.bfloat16
if t.dtype == torch.double:
return torch.float32
return t.dtype
if last_dtype is not None:
return last_dtype
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:
return last_tuple[1].dtype
for t in parameter.buffers():
last_dtype = t.dtype
if t.is_floating_point():
return t.dtype
return last_dtype
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):
"""
Returns the first found floating dtype in `state_dict` if there is one, otherwise returns the first dtype.
"""
for t in state_dict.values():
if t.is_floating_point():
return t.dtype
else:
return next(state_dict.values()).dtype
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
bit_search = re.search(r"[^\d](\d+)$", str(dtype))
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
def shard_checkpoint(
state_dict: Dict[str, torch.Tensor], max_shard_size: Union[int, str] = "10GB", weights_name: str = WEIGHTS_NAME
):
"""
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"`).
weights_name (`str`, *optional*, defaults to `"pytorch_model.bin"`):
The name of the model save file.
"""
max_shard_size = convert_file_size_to_int(max_shard_size)
'''
sharded_state_dicts = [{}]
last_block_size = 0
total_size = 0
storage_id_to_block = {}
for key, weight in state_dict.items():
storage_id = id_tensor_storage(weight)
# 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
weight_size = weight.numel() * dtype_byte_size(weight.dtype)
# If this weight is going to tip up over the maximal size, we split.
if last_block_size + weight_size > max_shard_size:
sharded_state_dicts.append({})
last_block_size = 0
sharded_state_dicts[-1][key] = weight
last_block_size += weight_size
total_size += weight_size
storage_id_to_block[storage_id] = len(sharded_state_dicts) - 1
'''
sharded_state_dicts = []
current_block = {}
current_block_size = 0
total_size = 0
for key, weight in state_dict.items():
weight_size = weight.numel() * dtype_byte_size(weight.dtype)
if current_block_size + weight_size > max_shard_size:
sharded_state_dicts.append(current_block)
current_block = {}
current_block_size = 0
current_block[key] = weight
current_block_size += weight_size
total_size += weight_size
sharded_state_dicts.append(current_block)
if len(sharded_state_dicts) == 1:
return {weights_name: sharded_state_dicts[0]}, None
weight_map = {}
shards = {}
for idx, shard in enumerate(sharded_state_dicts):
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"
)
shards[shard_file] = shard
for key in shard.keys():
weight_map[key] = shard_file
metadata = {"total_size": total_size}
index = {"metadata": metadata, "weight_map": weight_map}
return shards, index
def load_sharded_checkpoint(model, folder, strict=True, prefer_safe=True):
"""
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.
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.
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
"""
index_file = os.path.join(folder, WEIGHTS_INDEX_NAME)
safe_index_file = os.path.join(folder, SAFE_WEIGHTS_INDEX_NAME)
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
else:
logger.warning(
f"Cannot load sharded checkpoint at {folder} safely since safetensors is not installed!"
)
elif not index_present:
load_safe = True
load_index = safe_index_file if load_safe else index_file
with open(load_index, "r", encoding="utf-8") as f:
index = json.load(f)
shard_files = list(set(index["weight_map"].values()))
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)
loader = safe_load_file if load_safe else partial(torch.load, map_location="cpu")
for shard_file in shard_files:
state_dict = loader(os.path.join(folder, shard_file))
model.load_state_dict(state_dict, strict=False)
del state_dict
gc.collect()
return torch.nn.modules.module._IncompatibleKeys(missing_keys, unexpected_keys)
def load_state_dict(checkpoint_file: Union[str, os.PathLike]):
"""
Reads a PyTorch checkpoint file, returning properly formatted errors if they arise.
"""
if checkpoint_file.endswith(".safetensors") and is_safetensors_available():
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)
try:
return torch.load(checkpoint_file, map_location="cpu")
except Exception as e:
try:
with open(checkpoint_file) as f:
if f.read(7) == "version":
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."
)
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
def _load_state_dict_into_model(model_to_load, state_dict, start_prefix):
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)
metadata = getattr(state_dict, "_metadata", None)
state_dict = state_dict.copy()
if metadata is not None:
state_dict._metadata = metadata
error_msgs = []
def load(module: nn.Module, state_dict, prefix=""):
local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
args = (state_dict, prefix, local_metadata, True, [], [], error_msgs)
if len([key for key in state_dict if key.startswith(prefix)]) > 0:
if is_deepspeed_zero3_enabled():
import deepspeed
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:
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)
for name, child in module._modules.items():
if child is not None:
load(child, state_dict, prefix + name + ".")
load(model_to_load, state_dict, prefix=start_prefix)
del state_dict
return error_msgs
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`
"""
for k in loaded_state_dict_keys:
submodule, param_name = find_submodule_and_param_name(model, k, start_prefix)
if submodule is not None:
new_val = getattr(submodule, param_name)
if isinstance(new_val, torch.nn.Parameter):
new_val = torch.nn.Parameter(new_val.to("meta"))
else:
new_val = new_val.to("meta")
setattr(submodule, param_name, new_val)
def _load_state_dict_into_meta_model(
model,
state_dict,
loaded_state_dict_keys,
start_prefix,
expected_keys,
device_map=None,
offload_folder=None,
offload_index=None,
state_dict_folder=None,
state_dict_index=None,
dtype=None,
is_quantized=False,
is_safetensors=False,
keep_in_fp32_modules=None,
):
"""
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`
"""
if is_quantized:
from .utils.bitsandbytes import set_module_quantized_tensor_to_device
error_msgs = []
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)
for param_name, param in state_dict.items():
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
set_module_kwargs = {}
if dtype is not None and torch.is_floating_point(param):
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)
if "dtype" in list(inspect.signature(set_module_tensor_to_device).parameters):
set_module_kwargs["dtype"] = torch.float32
else:
param = param.to(dtype)
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)
set_module_kwargs["value"] = param
if device_map is None:
param_device = "cpu"
else:
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:
raise ValueError(f"{param_name} doesn't have any device set.")
param_device = device_map[module_name]
if param_device == "disk":
if not is_safetensors:
offload_index = offload_weight(param, param_name, offload_folder, offload_index)
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)
elif not is_quantized:
set_module_tensor_to_device(model, param_name, param_device, **set_module_kwargs)
else:
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:
set_module_quantized_tensor_to_device(
model, param_name, param_device, value=param, fp16_statistics=fp16_statistics
)
return error_msgs, offload_index, state_dict_index
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
class ModuleUtilsMixin:
"""
A few utilities for `torch.nn.Modules`, to be used as a mixin.
"""
@staticmethod
def _hook_rss_memory_pre_forward(module, *args, **kwargs):
try:
import psutil
except ImportError:
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
except ImportError:
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):
"""
Add a memory hook before and after each sub-module forward pass to record increase in memory consumption.
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()`.
"""
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):
"""
Reset the `mem_rss_diff` attribute of each module (see [`~modeling_utils.ModuleUtilsMixin.add_memory_hooks`]).
"""
for module in self.modules():
module.mem_rss_diff = 0
module.mem_rss_post_forward = 0
module.mem_rss_pre_forward = 0
@property
def device(self) -> torch.device:
"""
`torch.device`: The device on which the module is (assuming that all the module parameters are on the same
device).
"""
return get_parameter_device(self)
@property
def dtype(self) -> torch.dtype:
"""
`torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).
"""
return get_parameter_dtype(self)
def invert_attention_mask(self, encoder_attention_mask: Tensor) -> Tensor:
"""
Invert an attention mask (e.g., switches 0. and 1.).
Args:
encoder_attention_mask (`torch.Tensor`): An attention mask.
Returns:
`torch.Tensor`: The inverted attention mask.
"""
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, :]
encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=self.dtype)
encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * torch.finfo(self.dtype).min
return encoder_extended_attention_mask
@staticmethod
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
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]
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
def get_extended_attention_mask(
self, attention_mask: Tensor, input_shape: Tuple[int], device: torch.device = None,
dtype: torch.float = None
) -> Tensor:
"""
Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
Arguments:
attention_mask (`torch.Tensor`):
Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
input_shape (`Tuple[int]`):
The shape of the input to the model.
Returns:
`torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
"""
if dtype is None:
dtype = self.dtype
if not (attention_mask.dim() == 2 and self.config.is_decoder):
if device is not None:
warnings.warn(
"The `device` argument is deprecated and will be removed in v5 of Transformers.", FutureWarning
)
if attention_mask.dim() == 3:
extended_attention_mask = attention_mask[:, None, :, :]
elif attention_mask.dim() == 2:
if self.config.is_decoder:
extended_attention_mask = ModuleUtilsMixin.create_extended_attention_mask_for_decoder(
input_shape, attention_mask, device
)
else:
extended_attention_mask = attention_mask[:, None, None, :]
else:
raise ValueError(
f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"
)
extended_attention_mask = extended_attention_mask.to(dtype=dtype)
extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(dtype).min
return extended_attention_mask
def get_head_mask(
self, head_mask: Optional[Tensor], num_hidden_layers: int, is_attention_chunked: bool = False
) -> Tensor:
"""
Prepare the head mask if needed.
Args:
head_mask (`torch.Tensor` with shape `[num_heads]` or `[num_hidden_layers x num_heads]`, *optional*):
The mask indicating if we should keep the heads or not (1.0 for keep, 0.0 for discard).
num_hidden_layers (`int`):
The number of hidden layers in the model.
is_attention_chunked (`bool`, *optional*, defaults to `False`):
Whether or not the attentions scores are computed by chunks or not.
Returns:
`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.
"""
if head_mask is not None:
head_mask = self._convert_head_mask_to_5d(head_mask, num_hidden_layers)
if is_attention_chunked is True:
head_mask = head_mask.unsqueeze(-1)
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)
assert head_mask.dim() == 5, f"head_mask.dim != 5, instead {head_mask.dim()}"
head_mask = head_mask.to(dtype=self.dtype)
return head_mask
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:
only_trainable (`bool`, *optional*, defaults to `False`):
Whether or not to return only the number of trainable parameters
exclude_embeddings (`bool`, *optional*, defaults to `False`):
Whether or not to return only the number of non-embeddings parameters
Returns:
`int`: The number of parameters.
"""
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)
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:
inputs (`dict`): The model inputs.
Returns:
`int`: The total number of tokens.
"""
if not hasattr(self, "warnings_issued"):
self.warnings_issued = {}
if self.main_input_name in input_dict:
return input_dict[self.main_input_name].numel()
elif "estimate_tokens" not in self.warnings_issued:
logger.warning(
"Could not estimate the number of tokens of the input, floating-point operations will not be computed"
)
self.warnings_issued["estimate_tokens"] = True
return 0
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
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
re-use e.g. Albert or Universal Transformers, or if doing long-range modeling with very high sequence lengths.
Args:
batch_size (`int`):
The batch size for the forward pass.
sequence_length (`int`):
The number of tokens in each line of the batch.
exclude_embeddings (`bool`, *optional*, defaults to `True`):
Whether or not to count embedding and softmax operations.
Returns:
`int`: The number of floating-point operations.
"""
return 6 * self.estimate_tokens(input_dict) * self.num_parameters(exclude_embeddings=exclude_embeddings)
class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin):
r"""
Base class for all models.
[`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:
- resize the input embeddings,
- prune heads in the self-attention heads.
Class attributes (overridden by derived classes):
- **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:
- **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.
- **path** (`str`) -- A path to the TensorFlow checkpoint.
- **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.
- **is_parallelizable** (`bool`) -- A flag indicating whether this model supports model parallelization.
- **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).
"""
config_class = None
base_model_prefix = ""
main_input_name = "input_ids"
_auto_class = None
_no_split_modules = None
_skip_keys_device_placement = None
_keep_in_fp32_modules = None
_keys_to_ignore_on_load_missing = None
_keys_to_ignore_on_load_unexpected = None
_keys_to_ignore_on_save = None
is_parallelizable = False
supports_gradient_checkpointing = False
@property
def dummy_inputs(self) -> Dict[str, torch.Tensor]:
"""
`Dict[str, torch.Tensor]`: Dummy inputs to do a forward pass in the network.
"""
return {"input_ids": torch.tensor(DUMMY_INPUTS)}
@property
def framework(self) -> str:
"""
:str: Identifies that this is a PyTorch model.
"""
return "pt"
def __init__(self, config: PretrainedConfig, *inputs, **kwargs):
super().__init__()
if not isinstance(config, PretrainedConfig):
raise ValueError(
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)`"
)
self.config = config
self.name_or_path = config.name_or_path
self.warnings_issued = {}
self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
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()
delattr(self.config, "gradient_checkpointing")
@classmethod
def _from_config(cls, config, **kwargs):
"""
All context managers that the model should be initialized under go here.
Args:
torch_dtype (`torch.dtype`, *optional*):
Override the default `torch.dtype` and load the model under this dtype.
"""
torch_dtype = kwargs.pop("torch_dtype", None)
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")
with deepspeed.zero.Init(config_dict_or_path=deepspeed_config()):
model = cls(config, **kwargs)
else:
model = cls(config, **kwargs)
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:
dtype (`torch.dtype`):
a floating dtype to set to.
Returns:
`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`.
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.
"""
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
@property
def base_model(self) -> nn.Module:
"""
`torch.nn.Module`: The main body of the model.
"""
return getattr(self, self.base_model_prefix, self)
def can_generate(self) -> bool:
"""
Returns whether this model can generate sequences with `.generate()`.
Returns:
`bool`: Whether this model can generate sequences with `.generate()`.
"""
if "GenerationMixin" in str(self.prepare_inputs_for_generation.__func__):
return False
return True
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()
def get_input_embeddings(self) -> nn.Module:
"""
Returns the model's input embeddings.
Returns:
`nn.Module`: A torch module mapping vocabulary to hidden states.
"""
base_model = getattr(self, self.base_model_prefix, self)
if base_model is not self:
return base_model.get_input_embeddings()
else:
raise NotImplementedError
def set_input_embeddings(self, value: nn.Module):
"""
Set model's input embeddings.
Args:
value (`nn.Module`): A module mapping vocabulary to hidden states.
"""
base_model = getattr(self, self.base_model_prefix, self)
if base_model is not self:
base_model.set_input_embeddings(value)
else:
raise NotImplementedError
def get_output_embeddings(self) -> nn.Module:
"""
Returns the model's output embeddings.
Returns:
`nn.Module`: A torch module mapping hidden states to vocabulary.
"""
return None
def _init_weights(self, module):
"""
Initialize the weights. This method should be overridden by derived class.
"""
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
def tie_weights(self):
"""
Tie the weights between the input embeddings and the output embeddings.
If the `torchscript` flag is set in the configuration, can't handle parameter sharing so we are cloning the
weights instead.
"""
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())
if getattr(self.config, "is_encoder_decoder", False) and getattr(self.config, "tie_encoder_decoder", False):
if hasattr(self, self.base_model_prefix):
self = getattr(self, self.base_model_prefix)
self._tie_encoder_decoder_weights(self.encoder, self.decoder, self.base_model_prefix)
for module in self.modules():
if hasattr(module, "_tie_weights"):
module._tie_weights()
@staticmethod
def _tie_encoder_decoder_weights(encoder: nn.Module, decoder: nn.Module, base_model_prefix: str):
uninitialized_encoder_weights: List[str] = []
if decoder.__class__ != encoder.__class__:
logger.info(
f"{decoder.__class__} and {encoder.__class__} are not equal. In this case make sure that all encoder"
" weights are correctly initialized."
)
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
), f"{decoder_pointer} and {encoder_pointer} have to be of type nn.Module"
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}"
all_encoder_weights = {module_name + "/" + sub_name for sub_name in encoder_modules.keys()}
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
if not isinstance(decoder_modules[decoder_name], type(encoder_modules[encoder_name])) and len(
encoder_modules
) != len(decoder_modules):
encoder_layer_pos -= 1
continue
elif name not in encoder_modules:
continue
elif depth > 500:
raise ValueError(
"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."
)
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_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}"
)
def _tie_or_clone_weights(self, output_embeddings, input_embeddings):
"""Tie or clone module weights depending of whether we are using TorchScript or not"""
if self.config.torchscript:
output_embeddings.weight = nn.Parameter(input_embeddings.weight.clone())
else:
output_embeddings.weight = input_embeddings.weight
if getattr(output_embeddings, "bias", None) is not None:
output_embeddings.bias.data = nn.functional.pad(
output_embeddings.bias.data,
(
0,
output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0],
),
"constant",
0,
)
if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
output_embeddings.out_features = input_embeddings.num_embeddings
def resize_token_embeddings(self, new_num_tokens: Optional[int] = None) -> nn.Embedding:
"""
Resizes input token embeddings matrix of the model if `new_num_tokens != config.vocab_size`.
Takes care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
Arguments:
new_num_tokens (`int`, *optional*):
The number of new tokens in the embedding matrix. Increasing the size will add newly initialized
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.
Return:
`torch.nn.Embedding`: Pointer to the input tokens Embeddings Module of the model.
"""
model_embeds = self._resize_token_embeddings(new_num_tokens)
if new_num_tokens is None:
return model_embeds
self.config.vocab_size = new_num_tokens
self.vocab_size = new_num_tokens
self.tie_weights()
return model_embeds
def _resize_token_embeddings(self, new_num_tokens):
old_embeddings = self.get_input_embeddings()
new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
self.set_input_embeddings(new_embeddings)
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)
return self.get_input_embeddings()
def _get_resized_embeddings(
self, old_embeddings: nn.Embedding, new_num_tokens: Optional[int] = None
) -> nn.Embedding:
"""
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
Args:
old_embeddings (`torch.nn.Embedding`):
Old embeddings to be resized.
new_num_tokens (`int`, *optional*):
New number of tokens in the embedding matrix.
Increasing the size will add newly initialized 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.
Return:
`torch.nn.Embedding`: Pointer to the resized Embedding Module or the old Embedding Module if
`new_num_tokens` is `None`
"""
if new_num_tokens is None:
return old_embeddings
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()
if old_num_tokens == new_num_tokens:
return old_embeddings
if not isinstance(old_embeddings, nn.Embedding):
raise TypeError(
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}."
)
new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim)
new_embeddings.to(old_embeddings.weight.device, dtype=old_embeddings.weight.dtype)
self._init_weights(new_embeddings)
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, :]
return new_embeddings
def _get_resized_lm_head(
self, old_lm_head: nn.Linear, new_num_tokens: Optional[int] = None, transposed: Optional[bool] = False
) -> nn.Linear:
"""
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:
old_lm_head (`torch.nn.Linear`):
Old lm head liner layer to be resized.
new_num_tokens (`int`, *optional*):
New number of tokens in the linear matrix.
Increasing the size will add newly initialized 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.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`.
Return:
`torch.nn.Linear`: Pointer to the resized Linear Module or the old Linear Module if `new_num_tokens` is
`None`
"""
if new_num_tokens is None:
return old_lm_head
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()
)
if old_num_tokens == new_num_tokens:
return old_lm_head
if not isinstance(old_lm_head, nn.Linear):
raise TypeError(
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}."
)
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
new_lm_head = nn.Linear(*new_lm_head_shape, bias=has_new_lm_head_bias)
new_lm_head = new_lm_head.to(old_lm_head.weight.device, dtype=old_lm_head.weight.dtype)
self._init_weights(new_lm_head)
num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
if is_deepspeed_zero3_enabled():
import deepspeed
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):
if torch.distributed.get_rank() == 0:
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
]
if has_new_lm_head_bias:
new_lm_head.bias.data[:num_tokens_to_copy] = old_lm_head.bias.data[:num_tokens_to_copy]
else:
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]
if has_new_lm_head_bias:
new_lm_head.bias.data[:num_tokens_to_copy] = old_lm_head.bias.data[:num_tokens_to_copy]
return new_lm_head
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`"
)
def init_weights(self):
"""
If needed prunes and maybe initializes weights. If using a custom `PreTrainedModel`, you need to implement any
initialization logic in `_init_weights`.
"""
if self.config.pruned_heads:
self.prune_heads(self.config.pruned_heads)
if _init_weights:
self.apply(self._initialize_weights)
self.tie_weights()
def prune_heads(self, heads_to_prune: Dict[int, List[int]]):
"""
Prunes heads of the base model.
Arguments:
heads_to_prune (`Dict[int, List[int]]`):
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.
"""
for layer, heads in heads_to_prune.items():
union_heads = set(self.config.pruned_heads.get(layer, [])) | set(heads)
self.config.pruned_heads[layer] = list(union_heads)
self.base_model._prune_heads(heads_to_prune)
def gradient_checkpointing_enable(self):
"""
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))
def gradient_checkpointing_disable(self):
"""
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))
@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())
def save_pretrained(
self,
save_directory: Union[str, os.PathLike],
is_main_process: bool = True,
state_dict: Optional[dict] = None,
save_function: Callable = torch.save,
push_to_hub: bool = False,
max_shard_size: Union[int, str] = "10GB",
safe_serialization: bool = False,
variant: Optional[str] = None,
**kwargs,
):
"""
Save a model and its configuration file to a directory, so that it can be re-loaded using the
[`~PreTrainedModel.from_pretrained`] class method.
Arguments:
save_directory (`str` or `os.PathLike`):
Directory to which to save. Will be created if it doesn't exist.
is_main_process (`bool`, *optional*, defaults to `True`):
Whether the process calling this is the main process or not. Useful 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.
state_dict (nested dictionary of `torch.Tensor`):
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).
save_function (`Callable`):
The function to use to save the state dictionary. Useful on distributed training like TPUs when one
need to replace `torch.save` by another method.
push_to_hub (`bool`, *optional*, defaults to `False`):
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).
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>
safe_serialization (`bool`, *optional*, defaults to `False`):
Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
variant (`str`, *optional*):
If specified, weights are saved in the format pytorch_model.<variant>.bin.
kwargs:
Additional key word arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
"""
if getattr(self, "is_loaded_in_8bit", False) and getattr(self, "is_8bit_serializable", False):
warnings.warn(
"You are calling `save_pretrained` to a 8-bit converted model you may likely encounter unexepected"
" behaviors. If you want to save 8-bit models, make sure to have `bitsandbytes>0.37.2` installed.",
UserWarning,
)
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"
)
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")
if safe_serialization and not is_safetensors_available():
raise ImportError("`safe_serialization` requires the `safetensors library: `pip install safetensors`.")
if os.path.isfile(save_directory):
logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
return
os.makedirs(save_directory, exist_ok=True)
if push_to_hub:
commit_message = kwargs.pop("commit_message", None)
repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
repo_id = self._create_repo(repo_id, **kwargs)
files_timestamps = self._get_files_timestamps(save_directory)
model_to_save = unwrap_model(self)
dtype = get_parameter_dtype(model_to_save)
model_to_save.config.torch_dtype = str(dtype).split(".")[1]
model_to_save.config.architectures = [model_to_save.__class__.__name__]
if self._auto_class is not None:
custom_object_save(self, save_directory, config=self.config)
if is_main_process:
model_to_save.config.save_pretrained(save_directory)
if self.can_generate():
model_to_save.generation_config.save_pretrained(save_directory)
if state_dict is None:
state_dict = model_to_save.state_dict()
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)
if self._keys_to_ignore_on_save is not None:
for ignore_key in self._keys_to_ignore_on_save:
if ignore_key in state_dict.keys():
del state_dict[ignore_key]
if safe_serialization:
ptrs = collections.defaultdict(list)
for name, tensor in state_dict.items():
ident = (tensor.data_ptr(), tensor.device, tensor.shape, tensor.stride())
ptrs[ident].append(name)
shared_ptrs = {ptr: names for ptr, names in ptrs.items() if len(names) > 1}
warn_names = set()
for names in shared_ptrs.values():
if self._keys_to_ignore_on_load_missing is not None:
found = 0
for name in sorted(names):
matches_pattern = any(re.search(pat, name) for pat in self._keys_to_ignore_on_load_missing)
if matches_pattern and name in state_dict:
found += 1
if found < len(names):
del state_dict[name]
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",
)
weights_name = SAFE_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
weights_name = _add_variant(weights_name, variant)
shards, index = shard_checkpoint(state_dict, max_shard_size=max_shard_size, weights_name=weights_name)
for filename in os.listdir(save_directory):
full_filename = os.path.join(save_directory, filename)
weights_no_suffix = weights_name.replace(".bin", "").replace(".safetensors", "")
filename_no_suffix = filename.replace(".bin", "").replace(".safetensors", "")
reg = re.compile(r"(.*?)-\d{5}-of-\d{5}")
if (
filename.startswith(weights_no_suffix)
and os.path.isfile(full_filename)
and filename not in shards.keys()
and is_main_process
and reg.fullmatch(filename_no_suffix) is not None
):
os.remove(full_filename)
for shard_file, shard in shards.items():
if safe_serialization:
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))
if index is None:
path_to_weights = os.path.join(save_directory, _add_variant(WEIGHTS_NAME, variant))
logger.info(f"Model weights saved in {path_to_weights}")
else:
save_index_file = SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
save_index_file = os.path.join(save_directory, _add_variant(save_index_file, variant))
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}."
)
if push_to_hub:
self._upload_modified_files(
save_directory,
repo_id,
files_timestamps,
commit_message=commit_message,
token=kwargs.get("use_auth_token"),
)
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
def to(self, *args, **kwargs):
if getattr(self, "is_quantized", False):
raise ValueError(
"`.to` is not supported for `4-bit` or `8-bit` models. Please use the model as it is, since the"
" model has already been set to the correct devices and casted to the correct `dtype`."
)
else:
return super().to(*args, **kwargs)
def half(self, *args):
if getattr(self, "is_quantized", False):
raise ValueError(
"`.half()` is not supported for `4-bit` or `8-bit` models. Please use the model as it is, since the"
" model has already been casted to the correct `dtype`."
)
else:
return super().half(*args)
def float(self, *args):
if getattr(self, "is_quantized", False):
raise ValueError(
"`.float()` is not supported for `4-bit` or `8-bit` models. Please use the model as it is, since the"
" model has already been casted to the correct `dtype`."
)
else:
return super().float(*args)
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
r"""
Instantiate a pretrained pytorch model from a pre-trained model configuration.
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()`.
The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
task.
The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
weights are discarded.
Parameters:
pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
Can be either:
- A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
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`.
- 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
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.
- A path or url to a model folder containing a *flax checkpoint file* in *.msgpack* format (e.g,
`./flax_model/` containing `flax_model.msgpack`). In this case, `from_flax` should be set to
`True`.
- `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*):
Can be either:
- an instance of a class derived from [`PretrainedConfig`],
- a string or path valid as input to [`~PretrainedConfig.from_pretrained`].
Configuration for the model to use instead of an automatically loaded configuration. Configuration can
be automatically loaded when:
- The model is a model provided by the library (loaded with the *model id* string of a pretrained
model).
- The model was saved using [`~PreTrainedModel.save_pretrained`] and is reloaded by supplying the
save directory.
- 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*):
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
weights. In this case though, you should check if using [`~PreTrainedModel.save_pretrained`] and
[`~PreTrainedModel.from_pretrained`] is not a simpler option.
cache_dir (`Union[str, os.PathLike]`, *optional*):
Path to a directory in which a downloaded pretrained model configuration should be cached if the
standard cache should not be used.
from_tf (`bool`, *optional*, defaults to `False`):
Load the model weights from a TensorFlow checkpoint save file (see docstring of
`pretrained_model_name_or_path` argument).
from_flax (`bool`, *optional*, defaults to `False`):
Load the model weights from a Flax checkpoint save file (see docstring of
`pretrained_model_name_or_path` argument).
ignore_mismatched_sizes (`bool`, *optional*, defaults to `False`):
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).
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force the (re-)download of the model weights and configuration files, overriding the
cached versions if they exist.
resume_download (`bool`, *optional*, defaults to `False`):
Whether or not to delete incompletely received files. Will attempt to resume the download if such a
file exists.
proxies (`Dict[str, str]`, *optional*):
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.
output_loading_info(`bool`, *optional*, defaults to `False`):
Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
local_files_only(`bool`, *optional*, defaults to `False`):
Whether or not to only look at local files (i.e., do not try to download the model).
use_auth_token (`str` or `bool`, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, or not specified, will use
the token generated when running `huggingface-cli login` (stored in `~/.huggingface`).
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
identifier allowed by git.
<Tip>
To test a pull request you made on the Hub, you can pass `revision="refs/pr/<pr_number>".
</Tip>
mirror (`str`, *optional*):
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.
_fast_init(`bool`, *optional*, defaults to `True`):
Whether or not to disable fast initialization.
<Tip warning={true}>
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.
</Tip>
> Parameters for big model inference
low_cpu_mem_usage(`bool`, *optional*):
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*):
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>
device_map (`str` or `Dict[str, Union[int, str, torch.device]]` or `int` or `torch.device`, *optional*):
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
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.
To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
more information about each option see [designing a device
map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
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.
offload_folder (`str` or `os.PathLike`, *optional*):
If the `device_map` contains any value `"disk"`, the folder where we will offload weights.
offload_state_dict (`bool`, *optional*):
If `True`, will temporarily offload the CPU state dict to the hard drive to avoid getting out of CPU
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.
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.
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.
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.
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`.
kwargs (remaining dictionary of keyword arguments, *optional*):
Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
`output_attentions=True`). Behaves differently depending on whether a `config` is provided or
automatically loaded:
- 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
already been done)
- If a configuration is not provided, `kwargs` will be first passed to the configuration class
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.
<Tip>
Activate the special ["offline-mode"](https://huggingface.co/transformers/installation.html#offline-mode) to
use this method in a firewalled environment.
</Tip>
Examples:
```python
>>> from transformers import BertConfig, BertModel
>>> # Download model and configuration from huggingface.co and cache.
>>> model = BertModel.from_pretrained("bert-base-uncased")
>>> # Model was saved using *save_pretrained('./test/saved_model/')* (for example purposes, not runnable).
>>> model = BertModel.from_pretrained("./test/saved_model/")
>>> # Update configuration during loading.
>>> model = BertModel.from_pretrained("bert-base-uncased", output_attentions=True)
>>> assert model.config.output_attentions == True
>>> # Loading from a TF checkpoint file instead of a PyTorch model (slower, for example purposes, not runnable).
>>> 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)
>>> # Loading from a Flax checkpoint file instead of a PyTorch model (slower)
>>> model = BertModel.from_pretrained("bert-base-uncased", from_flax=True)
```
* `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
"""
config = kwargs.pop("config", None)
state_dict = kwargs.pop("state_dict", None)
cache_dir = kwargs.pop("cache_dir", None)
from_tf = kwargs.pop("from_tf", False)
from_flax = kwargs.pop("from_flax", False)
ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
force_download = kwargs.pop("force_download", False)
resume_download = kwargs.pop("resume_download", False)
proxies = kwargs.pop("proxies", None)
output_loading_info = kwargs.pop("output_loading_info", False)
local_files_only = kwargs.pop("local_files_only", False)
use_auth_token = kwargs.pop("use_auth_token", None)
revision = kwargs.pop("revision", None)
trust_remote_code = kwargs.pop("trust_remote_code", None)
_ = kwargs.pop("mirror", None)
from_pipeline = kwargs.pop("_from_pipeline", None)
from_auto_class = kwargs.pop("_from_auto", False)
_fast_init = kwargs.pop("_fast_init", True)
torch_dtype = kwargs.pop("torch_dtype", None)
low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", None)
device_map = kwargs.pop("device_map", None)
max_memory = kwargs.pop("max_memory", None)
offload_folder = kwargs.pop("offload_folder", None)
offload_state_dict = kwargs.pop("offload_state_dict", False)
load_in_8bit = kwargs.pop("load_in_8bit", False)
load_in_4bit = kwargs.pop("load_in_4bit", False)
quantization_config = kwargs.pop("quantization_config", None)
subfolder = kwargs.pop("subfolder", "")
commit_hash = kwargs.pop("_commit_hash", None)
variant = kwargs.pop("variant", None)
use_safetensors = kwargs.pop("use_safetensors", None if is_safetensors_available() else False)
if is_bitsandbytes_available():
is_8bit_serializable = version.parse(importlib_metadata.version("bitsandbytes")) > version.parse("0.37.2")
else:
is_8bit_serializable = False
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."
)
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}
if device_map is not None:
if low_cpu_mem_usage is None:
low_cpu_mem_usage = True
elif not low_cpu_mem_usage:
raise ValueError("Passing along a `device_map` requires `low_cpu_mem_usage=True`")
if low_cpu_mem_usage:
if device_map is not None:
require_version_core("torch>=1.10")
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`"
)
if quantization_config is None:
quantization_config, kwargs = BitsAndBytesConfig.from_dict(
config_dict={"load_in_8bit": load_in_8bit, "load_in_4bit": load_in_4bit},
return_unused_kwargs=True,
**kwargs,
)
elif quantization_config is not None:
load_in_8bit = quantization_config.load_in_8bit
load_in_4bit = quantization_config.load_in_4bit
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."
)
if load_in_8bit or load_in_4bit:
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` "
)
if torch_dtype is None:
logger.info(
f"Overriding torch_dtype={torch_dtype} with `torch_dtype=torch.float16` due to "
"requirements of `bitsandbytes` to enable model loading in 8-bit or 4-bit. "
"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."
)
torch_dtype = torch.float16
if device_map is None:
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
if from_tf or from_flax:
raise ValueError(
"Converting into 4-bit or 8-bit weights from tf/flax weights is currently not supported, please make"
" sure the weights are in PyTorch format."
)
from_pt = not (from_tf | from_flax)
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
if is_offline_mode() and not local_files_only:
logger.info("Offline mode: forcing local_files_only=True")
local_files_only = True
if not isinstance(config, PretrainedConfig):
config_path = config if config is not None else pretrained_model_name_or_path
config, model_kwargs = cls.config_class.from_pretrained(
config_path,
cache_dir=cache_dir,
return_unused_kwargs=True,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
subfolder=subfolder,
_from_auto=from_auto_class,
_from_pipeline=from_pipeline,
**kwargs,
)
else:
model_kwargs = kwargs
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:
if torch_dtype is None:
torch_dtype = torch.float16
if device_map is None:
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
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`."
)
if commit_hash is None:
commit_hash = getattr(config, "_commit_hash", None)
is_sharded = False
sharded_metadata = None
loading_info = None
keep_in_fp32_modules = None
use_keep_in_fp32_modules = False
if pretrained_model_name_or_path is not None:
pretrained_model_name_or_path = str(pretrained_model_name_or_path)
is_local = os.path.isdir(pretrained_model_name_or_path)
if is_local:
if from_tf and os.path.isfile(
os.path.join(pretrained_model_name_or_path, subfolder, TF_WEIGHTS_NAME + ".index")
):
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)
):
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)
):
archive_file = os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)
elif use_safetensors is not False and os.path.isfile(
os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant))
):
archive_file = os.path.join(
pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant)
)
elif use_safetensors is not False and os.path.isfile(
os.path.join(
pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
)
):
archive_file = os.path.join(
pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
)
is_sharded = True
elif os.path.isfile(
os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_NAME, variant))
):
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))
):
archive_file = os.path.join(
pretrained_model_name_or_path, subfolder, _add_variant(WEIGHTS_INDEX_NAME, variant)
)
is_sharded = True
elif os.path.isfile(
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)):
raise EnvironmentError(
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."
)
elif os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)):
raise EnvironmentError(
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."
)
else:
raise EnvironmentError(
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}."
)
elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path)):
archive_file = pretrained_model_name_or_path
is_local = True
elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path + ".index")):
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."
)
archive_file = os.path.join(subfolder, pretrained_model_name_or_path + ".index")
is_local = True
elif is_remote_url(pretrained_model_name_or_path):
filename = pretrained_model_name_or_path
resolved_archive_file = download_url(pretrained_model_name_or_path)
else:
if from_tf:
filename = TF2_WEIGHTS_NAME
elif from_flax:
filename = FLAX_WEIGHTS_NAME
elif use_safetensors is not False:
filename = _add_variant(SAFE_WEIGHTS_NAME, variant)
else:
filename = _add_variant(WEIGHTS_NAME, variant)
try:
cached_file_kwargs = {
"cache_dir": cache_dir,
"force_download": force_download,
"proxies": proxies,
"resume_download": resume_download,
"local_files_only": local_files_only,
"use_auth_token": use_auth_token,
"user_agent": user_agent,
"revision": revision,
"subfolder": subfolder,
"_raise_exceptions_for_missing_entries": False,
"_commit_hash": commit_hash,
}
resolved_archive_file = cached_file(pretrained_model_name_or_path, filename, **cached_file_kwargs)
if resolved_archive_file is None and filename == _add_variant(SAFE_WEIGHTS_NAME, variant):
resolved_archive_file = cached_file(
pretrained_model_name_or_path,
_add_variant(SAFE_WEIGHTS_INDEX_NAME, variant),
**cached_file_kwargs,
)
if resolved_archive_file is not None:
is_sharded = True
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`."
)
else:
filename = _add_variant(WEIGHTS_NAME, variant)
resolved_archive_file = cached_file(
pretrained_model_name_or_path, filename, **cached_file_kwargs
)
if resolved_archive_file is None and filename == _add_variant(WEIGHTS_NAME, variant):
resolved_archive_file = cached_file(
pretrained_model_name_or_path,
_add_variant(WEIGHTS_INDEX_NAME, variant),
**cached_file_kwargs,
)
if resolved_archive_file is not None:
is_sharded = True
if resolved_archive_file is None:
has_file_kwargs = {
"revision": revision,
"proxies": proxies,
"use_auth_token": use_auth_token,
}
if has_file(pretrained_model_name_or_path, TF2_WEIGHTS_NAME, **has_file_kwargs):
raise EnvironmentError(
f"{pretrained_model_name_or_path} does not appear to have a file named"
f" {_add_variant(WEIGHTS_NAME, variant)} but there is a file for TensorFlow weights."
" Use `from_tf=True` to load this model from those weights."
)
elif has_file(pretrained_model_name_or_path, FLAX_WEIGHTS_NAME, **has_file_kwargs):
raise EnvironmentError(
f"{pretrained_model_name_or_path} does not appear to have a file named"
f" {_add_variant(WEIGHTS_NAME, variant)} but there is a file for Flax weights. Use"
" `from_flax=True` to load this model from those weights."
)
elif variant is not None and has_file(
pretrained_model_name_or_path, WEIGHTS_NAME, **has_file_kwargs
):
raise EnvironmentError(
f"{pretrained_model_name_or_path} does not appear to have a file named"
f" {_add_variant(WEIGHTS_NAME, variant)} but there is a file without the variant"
f" {variant}. Use `variant=None` to load this model from those weights."
)
else:
raise EnvironmentError(
f"{pretrained_model_name_or_path} does not appear to have a file named"
f" {_add_variant(WEIGHTS_NAME, variant)}, {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME} or"
f" {FLAX_WEIGHTS_NAME}."
)
except EnvironmentError:
raise
except Exception:
raise EnvironmentError(
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"
f" directory containing a file named {_add_variant(WEIGHTS_NAME, variant)},"
f" {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME} or {FLAX_WEIGHTS_NAME}."
)
if is_local:
logger.info(f"loading weights file {archive_file}")
resolved_archive_file = archive_file
else:
logger.info(f"loading weights file {filename} from cache at {resolved_archive_file}")
else:
resolved_archive_file = None
if is_sharded:
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,
use_auth_token=use_auth_token,
user_agent=user_agent,
revision=revision,
subfolder=subfolder,
_commit_hash=commit_hash,
)
if from_pt:
if not is_sharded and state_dict is None:
state_dict = load_state_dict(resolved_archive_file)
dtype_orig = None
if torch_dtype is not None:
if isinstance(torch_dtype, str):
if torch_dtype == "auto":
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")
else:
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
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"
)
else:
raise ValueError(
f'`torch_dtype` can be either `torch.dtype` or `"auto"`, but received {torch_dtype}'
)
dtype_orig = cls._set_default_torch_dtype(torch_dtype)
use_keep_in_fp32_modules = (
(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)
)
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`"
)
if is_sharded:
loaded_state_dict_keys = sharded_metadata["all_checkpoint_keys"]
else:
loaded_state_dict_keys = list(state_dict.keys())
if low_cpu_mem_usage or use_keep_in_fp32_modules:
state_dict = None
config.name_or_path = pretrained_model_name_or_path
init_contexts = [no_init_weights(_enable=_fast_init)]
if is_deepspeed_zero3_enabled():
import deepspeed
logger.info("Detected DeepSpeed ZeRO-3: activating zero.init() for this model")
init_contexts = [deepspeed.zero.Init(config_dict_or_path=deepspeed_config())] + init_contexts
elif load_in_8bit or load_in_4bit or low_cpu_mem_usage:
init_contexts.append(init_empty_weights())
with ContextManagers(init_contexts):
model = cls(config, *model_args, **model_kwargs)
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 = []
if load_in_8bit or load_in_4bit:
from .utils.bitsandbytes import get_keys_to_not_convert, replace_with_bnb_linear
llm_int8_skip_modules = quantization_config.llm_int8_skip_modules
load_in_8bit_fp32_cpu_offload = quantization_config.llm_int8_enable_fp32_cpu_offload
logger.info("Detected 8-bit loading: activating 8-bit loading for this model")
if llm_int8_skip_modules is None:
modules_to_not_convert = get_keys_to_not_convert(model)
else:
modules_to_not_convert = llm_int8_skip_modules
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)
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 "
"`llm_int8_enable_fp32_cpu_offload=True`. Note that these modules will not be "
" converted to 8-bit but kept in 32-bit."
)
modules_to_not_convert.extend(keys_on_cpu)
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
)
model._is_quantized_training_enabled = version.parse(
importlib_metadata.version("bitsandbytes")
) >= version.parse("0.37.0")
model.config.quantization_config = quantization_config
model.is_8bit_serializable = is_8bit_serializable
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."
"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."
)
if isinstance(device_map, str):
special_dtypes = {}
if load_in_8bit or load_in_4bit:
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)
}
)
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,"
"`pip install --upgrade accelerate` or install it from source to support fp4 auto device map"
"calculation. You may encounter unexpected behavior, or pass your own device map"
)
elif load_in_8bit:
target_dtype = torch.int8
if model._no_split_modules is None:
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."
)
no_split_modules = model._no_split_modules
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'."
)
elif device_map in ["balanced", "balanced_low_0"] and get_balanced_memory is None:
raise ValueError(f"`device_map={device_map}` requires a source install of Accelerate.")
kwargs = {"no_split_module_classes": no_split_modules}
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`)."
)
if device_map != "sequential" and get_balanced_memory is not None:
max_memory = get_balanced_memory(
model,
dtype=target_dtype,
low_zero=(device_map == "balanced_low_0"),
max_memory=max_memory,
**kwargs,
)
kwargs["max_memory"] = max_memory
model.tie_weights()
device_map = infer_auto_device_map(model, dtype=target_dtype, **kwargs)
if load_in_8bit or load_in_4bit:
device_map_without_lm_head = {
key: device_map[key] for key in device_map.keys() if key not in modules_to_not_convert
}
if "cpu" in device_map_without_lm_head.values() or "disk" in device_map_without_lm_head.values():
raise ValueError(
"""
Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit
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.
"""
)
del device_map_without_lm_head
elif device_map is not None:
model.tie_weights()
tied_params = find_tied_parameters(model)
if check_tied_parameters_on_same_device is not None:
check_tied_parameters_on_same_device(tied_params, device_map)
if from_tf:
if resolved_archive_file.endswith(".index"):
model = cls.load_tf_weights(model, config, resolved_archive_file[:-6])
else:
try:
from .modeling_tf_pytorch_utils import load_tf2_checkpoint_in_pytorch_model
model, loading_info = load_tf2_checkpoint_in_pytorch_model(
model, resolved_archive_file, allow_missing_keys=True, output_loading_info=True
)
except ImportError:
logger.error(
"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."
)
raise
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(
"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."
)
raise
elif from_pt:
if dtype_orig is not None:
torch.set_default_dtype(dtype_orig)
(
model,
missing_keys,
unexpected_keys,
mismatched_keys,
offload_index,
error_msgs,
) = cls._load_pretrained_model(
model,
state_dict,
loaded_state_dict_keys,
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,
device_map=device_map,
offload_folder=offload_folder,
offload_state_dict=offload_state_dict,
dtype=torch_dtype,
is_quantized=(load_in_8bit or load_in_4bit),
keep_in_fp32_modules=keep_in_fp32_modules,
)
model.is_loaded_in_4bit = load_in_4bit
model.is_loaded_in_8bit = load_in_8bit
model.is_quantized = load_in_8bit or load_in_4bit
model.tie_weights()
model.eval()
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,
use_auth_token=use_auth_token,
revision=revision,
subfolder=subfolder,
_from_auto=from_auto_class,
_from_pipeline=from_pipeline,
**kwargs,
)
except (OSError, TypeError):
logger.info(
"Generation config file not found, using a generation config created from the model config."
)
pass
if device_map is not None:
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)
if output_loading_info:
if loading_info is None:
loading_info = {
"missing_keys": missing_keys,
"unexpected_keys": unexpected_keys,
"mismatched_keys": mismatched_keys,
"error_msgs": error_msgs,
}
return model, loading_info
return model
@classmethod
def _load_pretrained_model(
cls,
model,
state_dict,
loaded_keys,
resolved_archive_file,
pretrained_model_name_or_path,
ignore_mismatched_sizes=False,
sharded_metadata=None,
_fast_init=True,
low_cpu_mem_usage=False,
device_map=None,
offload_folder=None,
offload_state_dict=None,
dtype=None,
is_quantized=False,
keep_in_fp32_modules=None,
):
is_safetensors = False
if is_quantized:
from .utils.bitsandbytes import set_module_quantized_tensor_to_device
if device_map is not None and "disk" in device_map.values():
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:
raise ValueError(
"The current `device_map` had weights offloaded to the disk. Please provide an `offload_folder`"
" for them. Alternatively, make sure you have `safetensors` installed if the model you are using"
" offers the weights in this format."
)
if offload_folder is not None:
os.makedirs(offload_folder, exist_ok=True)
if offload_state_dict is None:
offload_state_dict = True
is_sharded_safetensors = is_safetensors and sharded_metadata is not None
model_state_dict = model.state_dict()
expected_keys = list(model_state_dict.keys())
prefix = model.base_model_prefix
def _fix_key(key):
if "beta" in key:
return key.replace("beta", "bias")
if "gamma" in key:
return key.replace("gamma", "weight")
return key
original_loaded_keys = loaded_keys
loaded_keys = [_fix_key(key) for key in loaded_keys]
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
remove_prefix_from_model = not has_prefix_module and expects_prefix_module
add_prefix_to_model = has_prefix_module and not expects_prefix_module
if remove_prefix_from_model:
_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]
elif add_prefix_to_model:
expected_keys = [".".join([prefix, s]) for s in expected_keys]
missing_keys = list(set(expected_keys) - set(loaded_keys))
unexpected_keys = list(set(loaded_keys) - set(expected_keys))
if find_tied_parameters is not None:
model.tie_weights()
tied_params = find_tied_parameters(model)
else:
tied_params = []
_missing = []
for k in missing_keys:
found = False
for group in tied_params:
if k in group:
found = True
if len(group) > 2:
group.remove(k)
else:
_missing.append(k)
if not found:
_missing.append(k)
missing_keys = _missing
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]
if low_cpu_mem_usage:
for key in missing_keys:
if key in list(model_state_dict.keys()):
pass
elif f"{prefix}.{key}" in list(model_state_dict.keys()):
key = f"{prefix}.{key}"
elif key.startswith(prefix) and ".".join(key.split(".")[1:]) in list(model_state_dict.keys()):
key = ".".join(key.split(".")[1:])
param = model_state_dict[key]
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
if param.device == torch.device("meta"):
if not (is_quantized):
set_module_tensor_to_device(model, key, "cpu", torch.empty(*param.size(), dtype=target_dtype))
else:
set_module_quantized_tensor_to_device(
model, key, "cpu", torch.empty(*param.size(), dtype=target_dtype)
)
if _fast_init:
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)
model.apply(model._initialize_weights)
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)
start_prefix = ""
model_to_load = model
if len(cls.base_model_prefix) > 0 and not hasattr(model, cls.base_model_prefix) and has_prefix_module:
start_prefix = cls.base_model_prefix + "."
if len(cls.base_model_prefix) > 0 and hasattr(model, cls.base_model_prefix) and not has_prefix_module:
model_to_load = getattr(model, cls.base_model_prefix)
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):
raise ValueError(
"The state dictionary of the model you are trying to load is corrupted. Are you sure it was "
"properly saved?"
)
if device_map is not None:
device_map = {k.replace(f"{cls.base_model_prefix}.", ""): v for k, v in device_map.items()}
def _find_mismatched_keys(
state_dict,
model_state_dict,
loaded_keys,
add_prefix_to_model,
remove_prefix_from_model,
ignore_mismatched_sizes,
):
mismatched_keys = []
if ignore_mismatched_sizes:
for checkpoint_key in loaded_keys:
model_key = checkpoint_key
if remove_prefix_from_model:
model_key = f"{prefix}.{checkpoint_key}"
elif add_prefix_to_model:
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]
return mismatched_keys
if resolved_archive_file is not None:
folder = os.path.sep.join(resolved_archive_file[0].split(os.path.sep)[:-1])
else:
folder = None
if device_map is not None and is_safetensors:
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()}
offload_index = {
p: {"safetensors_file": f, "weight_name": p, "dtype": str_dtype}
for p, f in weight_map.items()
if param_device_map[p] == "disk"
}
if state_dict is not None:
mismatched_keys = _find_mismatched_keys(
state_dict,
model_state_dict,
original_loaded_keys,
add_prefix_to_model,
remove_prefix_from_model,
ignore_mismatched_sizes,
)
error_msgs = _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
offload_index = None
else:
if not isinstance(resolved_archive_file, list):
resolved_archive_file = [resolved_archive_file]
error_msgs = []
mismatched_keys = []
if not is_safetensors:
offload_index = {} if device_map is not None and "disk" in device_map.values() else None
if offload_state_dict:
state_dict_folder = tempfile.mkdtemp()
state_dict_index = {}
else:
state_dict_folder = None
state_dict_index = None
if is_sharded_safetensors:
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 = []
if len(resolved_archive_file) > 1:
resolved_archive_file = logging.tqdm(resolved_archive_file, desc="Loading checkpoint shards")
for shard_file in resolved_archive_file:
if shard_file in disk_only_shard_files:
continue
state_dict = load_state_dict(shard_file)
mismatched_keys += _find_mismatched_keys(
state_dict,
model_state_dict,
original_loaded_keys,
add_prefix_to_model,
remove_prefix_from_model,
ignore_mismatched_sizes,
)
if low_cpu_mem_usage:
new_error_msgs, offload_index, state_dict_index = _load_state_dict_into_meta_model(
model_to_load,
state_dict,
loaded_keys,
start_prefix,
expected_keys,
device_map=device_map,
offload_folder=offload_folder,
offload_index=offload_index,
state_dict_folder=state_dict_folder,
state_dict_index=state_dict_index,
dtype=dtype,
is_quantized=is_quantized,
is_safetensors=is_safetensors,
keep_in_fp32_modules=keep_in_fp32_modules,
)
error_msgs += new_error_msgs
else:
error_msgs += _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
del state_dict
gc.collect()
if offload_index is not None and len(offload_index) > 0:
if model != model_to_load:
prefix = cls.base_model_prefix
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"),
)
offload_index = {f"{prefix}.{key}": value for key, value in offload_index.items()}
if not is_safetensors:
save_offload_index(offload_index, offload_folder)
offload_index = None
if offload_state_dict:
load_offloaded_weights(model_to_load, state_dict_index, state_dict_folder)
shutil.rmtree(state_dict_folder)
if len(error_msgs) > 0:
error_msg = "\n\t".join(error_msgs)
if "size mismatch" in error_msg:
error_msg += (
"\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
)
raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")
if is_quantized:
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]
if len(unexpected_keys) > 0:
logger.warning(
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)."
)
else:
logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
if len(missing_keys) > 0:
logger.warning(
f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"
" TRAIN this model on a down-stream task to be able to use it for predictions and inference."
)
elif len(mismatched_keys) == 0:
logger.info(
f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"
f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the checkpoint"
f" was trained on, you can already use {model.__class__.__name__} for predictions without further"
" training."
)
if len(mismatched_keys) > 0:
mismatched_warning = "\n".join(
[
f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated"
for key, shape1, shape2 in mismatched_keys
]
)
logger.warning(
f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not"
f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be able"
" to use it for predictions and inference."
)
return model, missing_keys, unexpected_keys, mismatched_keys, offload_index, error_msgs
def retrieve_modules_from_names(self, names, add_prefix=False, remove_prefix=False):
module_keys = {".".join(key.split(".")[:-1]) for key in names}
module_keys = module_keys.union(
{".".join(key.split(".")[:-2]) for key in names if len(key) > 0 and key[-1].isdigit()}
)
retrieved_modules = []
for name, module in self.named_modules():
if remove_prefix:
_prefix = f"{self.base_model_prefix}."
name = name[len(_prefix):] if name.startswith(_prefix) else name
elif add_prefix:
name = ".".join([self.base_model_prefix, name]) if len(name) > 0 else self.base_model_prefix
if name in module_keys:
retrieved_modules.append(module)
return retrieved_modules
@staticmethod
def _load_pretrained_model_low_mem(model, loaded_state_dict_keys, resolved_archive_file, start_prefix=""):
"""
This is an experimental function that loads the model using ~1.x model size CPU memory
Before you call it do:
1. save which state_dict keys are available
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.
"""
_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
@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.
<Tip warning={true}>
This API is experimental and may have some slight breaking changes in the next releases.
</Tip>
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
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)
PreTrainedModel.push_to_hub = copy_func(PreTrainedModel.push_to_hub)
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"
)
class PoolerStartLogits(nn.Module):
"""
Compute SQuAD start logits from sequence hidden states.
Args:
config ([`PretrainedConfig`]):
The config used by the model, will be used to grab the `hidden_size` of the model.
"""
def __init__(self, config: PretrainedConfig):
super().__init__()
self.dense = nn.Linear(config.hidden_size, 1)
def forward(
self, hidden_states: torch.FloatTensor, p_mask: Optional[torch.FloatTensor] = None
) -> torch.FloatTensor:
"""
Args:
hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
The final hidden states of the model.
p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
should be masked.
Returns:
`torch.FloatTensor`: The start logits for SQuAD.
"""
x = self.dense(hidden_states).squeeze(-1)
if p_mask is not None:
if get_parameter_dtype(self) == torch.float16:
x = x * (1 - p_mask) - 65500 * p_mask
else:
x = x * (1 - p_mask) - 1e30 * p_mask
return x
class PoolerEndLogits(nn.Module):
"""
Compute SQuAD end logits from sequence hidden states.
Args:
config ([`PretrainedConfig`]):
The config used by the model, will be used to grab the `hidden_size` of the model and the `layer_norm_eps`
to use.
"""
def __init__(self, config: PretrainedConfig):
super().__init__()
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)
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:
hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
The final hidden states of the model.
start_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
The hidden states of the first tokens for the labeled span.
start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
The position of the first token for the labeled span.
p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
should be masked.
<Tip>
One of `start_states` or `start_positions` should be not `None`. If both are set, `start_positions` overrides
`start_states`.
</Tip>
Returns:
`torch.FloatTensor`: The end logits for SQuAD.
"""
assert (
start_states is not None or start_positions is not None
), "One of start_states, start_positions should be not None"
if start_positions is not None:
slen, hsz = hidden_states.shape[-2:]
start_positions = start_positions[:, None, None].expand(-1, -1, hsz)
start_states = hidden_states.gather(-2, start_positions)
start_states = start_states.expand(-1, slen, -1)
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:
if get_parameter_dtype(self) == torch.float16:
x = x * (1 - p_mask) - 65500 * p_mask
else:
x = x * (1 - p_mask) - 1e30 * p_mask
return x
class PoolerAnswerClass(nn.Module):
"""
Compute SQuAD 2.0 answer class from classification and start tokens hidden states.
Args:
config ([`PretrainedConfig`]):
The config used by the model, will be used to grab the `hidden_size` of the model.
"""
def __init__(self, config):
super().__init__()
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)
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:
"""
Args:
hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
The final hidden states of the model.
start_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
The hidden states of the first tokens for the labeled span.
start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
The position of the first token for the labeled span.
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>
One of `start_states` or `start_positions` should be not `None`. If both are set, `start_positions` overrides
`start_states`.
</Tip>
Returns:
`torch.FloatTensor`: The SQuAD 2.0 answer class.
"""
hsz = hidden_states.shape[-1]
assert (
start_states is not None or start_positions is not None
), "One of start_states, start_positions should be not None"
if start_positions is not None:
start_positions = start_positions[:, None, None].expand(-1, -1, hsz)
start_states = hidden_states.gather(-2, start_positions).squeeze(-2)
if cls_index is not None:
cls_index = cls_index[:, None, None].expand(-1, -1, hsz)
cls_token_state = hidden_states.gather(-2, cls_index).squeeze(-2)
else:
cls_token_state = hidden_states[:, -1, :]
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
@dataclass
class SquadHeadOutput(ModelOutput):
"""
Base class for outputs of question answering models using a [`~modeling_utils.SQuADHead`].
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned if both `start_positions` and `end_positions` are provided):
Classification loss as the sum of start token, end token (and is_impossible if provided) classification
losses.
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):
Log probabilities for the top config.start_n_top start token possibilities (beam-search).
start_top_index (`torch.LongTensor` of shape `(batch_size, config.start_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
Indices for the top config.start_n_top start token possibilities (beam-search).
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
(beam-search).
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.
"""
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
class SQuADHead(nn.Module):
r"""
A SQuAD head inspired by XLNet.
Args:
config ([`PretrainedConfig`]):
The config used by the model, will be used to grab the `hidden_size` of the model and the `layer_norm_eps`
to use.
"""
def __init__(self, config):
super().__init__()
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)
@replace_return_docstrings(output_type=SquadHeadOutput, config_class=PretrainedConfig)
def forward(
self,
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,
return_dict: bool = False,
) -> Union[SquadHeadOutput, Tuple[torch.FloatTensor]]:
"""
Args:
hidden_states (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`):
Final hidden states of the model on the sequence tokens.
start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Positions of the first token for the labeled span.
end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Positions of the last token for the labeled span.
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*):
Whether the question has a possible answer in the paragraph or not.
p_mask (`torch.FloatTensor` of shape `(batch_size, seq_len)`, *optional*):
Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
should be masked.
return_dict (`bool`, *optional*, defaults to `False`):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
Returns:
"""
start_logits = self.start_logits(hidden_states, p_mask=p_mask)
if start_positions is not None and end_positions is not None:
for x in (start_positions, end_positions, cls_index, is_impossible):
if x is not None and x.dim() > 1:
x.squeeze_(-1)
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:
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)
total_loss += cls_loss * 0.5
return SquadHeadOutput(loss=total_loss) if return_dict else (total_loss,)
else:
bsz, slen, hsz = hidden_states.size()
start_log_probs = nn.functional.softmax(start_logits, dim=-1)
start_top_log_probs, start_top_index = torch.topk(
start_log_probs, self.start_n_top, dim=-1
)
start_top_index_exp = start_top_index.unsqueeze(-1).expand(-1, -1, hsz)
start_states = torch.gather(hidden_states, -2, start_top_index_exp)
start_states = start_states.unsqueeze(1).expand(-1, slen, -1, -1)
hidden_states_expanded = hidden_states.unsqueeze(2).expand_as(
start_states
)
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)
end_log_probs = nn.functional.softmax(end_logits, dim=1)
end_top_log_probs, end_top_index = torch.topk(
end_log_probs, self.end_n_top, dim=1
)
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)
if not return_dict:
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,
)
class SequenceSummary(nn.Module):
r"""
Compute a single vector summary of a sequence hidden states.
Args:
config ([`PretrainedConfig`]):
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):
- **summary_type** (`str`) -- The method to use to make this summary. Accepted values are:
- `"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
- **summary_use_proj** (`bool`) -- Add a projection after the vector extraction.
- **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.
"""
def __init__(self, config: PretrainedConfig):
super().__init__()
self.summary_type = getattr(config, "summary_type", "last")
if self.summary_type == "attn":
raise NotImplementedError
self.summary = Identity()
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:
num_classes = config.num_labels
else:
num_classes = config.hidden_size
self.summary = nn.Linear(config.hidden_size, num_classes)
activation_string = getattr(config, "summary_activation", None)
self.activation: Callable = get_activation(activation_string) if activation_string else Identity()
self.first_dropout = Identity()
if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
self.first_dropout = nn.Dropout(config.summary_first_dropout)
self.last_dropout = Identity()
if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
self.last_dropout = nn.Dropout(config.summary_last_dropout)
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:
hidden_states (`torch.FloatTensor` of shape `[batch_size, seq_len, hidden_size]`):
The hidden states of the last layer.
cls_index (`torch.LongTensor` of shape `[batch_size]` or `[batch_size, ...]` where ... are optional leading dimensions of `hidden_states`, *optional*):
Used if `summary_type == "cls_index"` and takes the last token of the sequence as classification token.
Returns:
`torch.FloatTensor`: The summary of the sequence hidden states.
"""
if self.summary_type == "last":
output = hidden_states[:, -1]
elif self.summary_type == "first":
output = hidden_states[:, 0]
elif self.summary_type == "mean":
output = hidden_states.mean(dim=1)
elif self.summary_type == "cls_index":
if cls_index is None:
cls_index = torch.full_like(
hidden_states[..., :1, :],
hidden_states.shape[-2] - 1,
dtype=torch.long,
)
else:
cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
output = hidden_states.gather(-2, cls_index).squeeze(-2)
elif self.summary_type == "attn":
raise NotImplementedError
output = self.first_dropout(output)
output = self.summary(output)
output = self.activation(output)
output = self.last_dropout(output)
return output
def unwrap_model(model: nn.Module) -> nn.Module:
"""
Recursively unwraps a model from potential containers (as used in distributed training).
Args:
model (`torch.nn.Module`): The model to unwrap.
"""
if hasattr(model, "module"):
return unwrap_model(model.module)
else:
return model
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"}]