File size: 17,136 Bytes
9c6594c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 |
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Extends `dill` to support pickling more types and produce more consistent dumps."""
import os
import sys
from io import BytesIO
from types import CodeType, FunctionType
import dill
from packaging import version
from .. import config
class Pickler(dill.Pickler):
dispatch = dill._dill.MetaCatchingDict(dill.Pickler.dispatch.copy())
_legacy_no_dict_keys_sorting = False
def save(self, obj, save_persistent_id=True):
obj_type = type(obj)
if obj_type not in self.dispatch:
if "regex" in sys.modules:
import regex # type: ignore
if obj_type is regex.Pattern:
pklregister(obj_type)(_save_regexPattern)
if "spacy" in sys.modules:
import spacy # type: ignore
if issubclass(obj_type, spacy.Language):
pklregister(obj_type)(_save_spacyLanguage)
if "tiktoken" in sys.modules:
import tiktoken # type: ignore
if obj_type is tiktoken.Encoding:
pklregister(obj_type)(_save_tiktokenEncoding)
if "torch" in sys.modules:
import torch # type: ignore
if issubclass(obj_type, torch.Tensor):
pklregister(obj_type)(_save_torchTensor)
if obj_type is torch.Generator:
pklregister(obj_type)(_save_torchGenerator)
# Unwrap `torch.compile`-ed modules
if issubclass(obj_type, torch.nn.Module):
obj = getattr(obj, "_orig_mod", obj)
if "transformers" in sys.modules:
import transformers # type: ignore
if issubclass(obj_type, transformers.PreTrainedTokenizerBase):
pklregister(obj_type)(_save_transformersPreTrainedTokenizerBase)
# Unwrap `torch.compile`-ed functions
if obj_type is FunctionType:
obj = getattr(obj, "_torchdynamo_orig_callable", obj)
dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
def _batch_setitems(self, items):
if self._legacy_no_dict_keys_sorting:
return super()._batch_setitems(items)
# Ignore the order of keys in a dict
try:
# Faster, but fails for unorderable elements
items = sorted(items)
except Exception: # TypeError, decimal.InvalidOperation, etc.
from datasets.fingerprint import Hasher
items = sorted(items, key=lambda x: Hasher.hash(x[0]))
dill.Pickler._batch_setitems(self, items)
def memoize(self, obj):
# Don't memoize strings since two identical strings can have different Python ids
if type(obj) is not str: # noqa: E721
dill.Pickler.memoize(self, obj)
def pklregister(t):
"""Register a custom reducer for the type."""
def proxy(func):
Pickler.dispatch[t] = func
return func
return proxy
def dump(obj, file):
"""Pickle an object to a file."""
Pickler(file, recurse=True).dump(obj)
def dumps(obj):
"""Pickle an object to a string."""
file = BytesIO()
dump(obj, file)
return file.getvalue()
if config.DILL_VERSION < version.parse("0.3.6"):
def log(pickler, msg):
dill._dill.log.info(msg)
elif config.DILL_VERSION.release[:3] in [
version.parse("0.3.6").release,
version.parse("0.3.7").release,
version.parse("0.3.8").release,
]:
def log(pickler, msg):
dill._dill.logger.trace(pickler, msg)
@pklregister(set)
def _save_set(pickler, obj):
log(pickler, f"Se: {obj}")
try:
# Faster, but fails for unorderable elements
args = (sorted(obj),)
except Exception: # TypeError, decimal.InvalidOperation, etc.
from datasets.fingerprint import Hasher
args = (sorted(obj, key=Hasher.hash),)
pickler.save_reduce(set, args, obj=obj)
log(pickler, "# Se")
def _save_regexPattern(pickler, obj):
import regex # type: ignore
log(pickler, f"Re: {obj}")
args = (obj.pattern, obj.flags)
pickler.save_reduce(regex.compile, args, obj=obj)
log(pickler, "# Re")
def _save_tiktokenEncoding(pickler, obj):
import tiktoken # type: ignore
log(pickler, f"Enc: {obj}")
args = (obj.name, obj._pat_str, obj._mergeable_ranks, obj._special_tokens)
pickler.save_reduce(tiktoken.Encoding, args, obj=obj)
log(pickler, "# Enc")
def _save_torchTensor(pickler, obj):
import torch # type: ignore
# `torch.from_numpy` is not picklable in `torch>=1.11.0`
def create_torchTensor(np_array, dtype=None):
tensor = torch.from_numpy(np_array)
if dtype:
tensor = tensor.type(dtype)
return tensor
log(pickler, f"To: {obj}")
if obj.dtype == torch.bfloat16:
args = (obj.detach().to(torch.float).cpu().numpy(), torch.bfloat16)
else:
args = (obj.detach().cpu().numpy(),)
pickler.save_reduce(create_torchTensor, args, obj=obj)
log(pickler, "# To")
def _save_torchGenerator(pickler, obj):
import torch # type: ignore
def create_torchGenerator(state):
generator = torch.Generator()
generator.set_state(state)
return generator
log(pickler, f"Ge: {obj}")
args = (obj.get_state(),)
pickler.save_reduce(create_torchGenerator, args, obj=obj)
log(pickler, "# Ge")
def _save_spacyLanguage(pickler, obj):
import spacy # type: ignore
def create_spacyLanguage(config, bytes):
lang_cls = spacy.util.get_lang_class(config["nlp"]["lang"])
lang_inst = lang_cls.from_config(config)
return lang_inst.from_bytes(bytes)
log(pickler, f"Sp: {obj}")
args = (obj.config, obj.to_bytes())
pickler.save_reduce(create_spacyLanguage, args, obj=obj)
log(pickler, "# Sp")
def _save_transformersPreTrainedTokenizerBase(pickler, obj):
log(pickler, f"Tok: {obj}")
# Ignore the `cache` attribute
state = obj.__dict__
if "cache" in state and isinstance(state["cache"], dict):
state["cache"] = {}
pickler.save_reduce(type(obj), (), state=state, obj=obj)
log(pickler, "# Tok")
if config.DILL_VERSION < version.parse("0.3.6"):
@pklregister(CodeType)
def _save_code(pickler, obj):
"""
From dill._dill.save_code
This is a modified version that removes the origin (filename + line no.)
of functions created in notebooks or shells for example.
"""
dill._dill.log.info(f"Co: {obj}")
# The filename of a function is the .py file where it is defined.
# Filenames of functions created in notebooks or shells start with '<'
# ex: <ipython-input-13-9ed2afe61d25> for ipython, and <stdin> for shell
# Filenames of functions created in ipykernel the filename
# look like f"{tempdir}/ipykernel_{id1}/{id2}.py"
# Moreover lambda functions have a special name: '<lambda>'
# ex: (lambda x: x).__code__.co_name == "<lambda>" # True
#
# For the hashing mechanism we ignore where the function has been defined
# More specifically:
# - we ignore the filename of special functions (filename starts with '<')
# - we always ignore the line number
# - we only use the base name of the file instead of the whole path,
# to be robust in case a script is moved for example.
#
# Only those two lines are different from the original implementation:
co_filename = (
""
if obj.co_filename.startswith("<")
or (
len(obj.co_filename.split(os.path.sep)) > 1
and obj.co_filename.split(os.path.sep)[-2].startswith("ipykernel_")
)
or obj.co_name == "<lambda>"
else os.path.basename(obj.co_filename)
)
co_firstlineno = 1
# The rest is the same as in the original dill implementation
if dill._dill.PY3:
if hasattr(obj, "co_posonlyargcount"):
args = (
obj.co_argcount,
obj.co_posonlyargcount,
obj.co_kwonlyargcount,
obj.co_nlocals,
obj.co_stacksize,
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_filename,
obj.co_name,
co_firstlineno,
obj.co_lnotab,
obj.co_freevars,
obj.co_cellvars,
)
else:
args = (
obj.co_argcount,
obj.co_kwonlyargcount,
obj.co_nlocals,
obj.co_stacksize,
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_filename,
obj.co_name,
co_firstlineno,
obj.co_lnotab,
obj.co_freevars,
obj.co_cellvars,
)
else:
args = (
obj.co_argcount,
obj.co_nlocals,
obj.co_stacksize,
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_filename,
obj.co_name,
co_firstlineno,
obj.co_lnotab,
obj.co_freevars,
obj.co_cellvars,
)
pickler.save_reduce(CodeType, args, obj=obj)
dill._dill.log.info("# Co")
return
elif config.DILL_VERSION.release[:3] in [
version.parse("0.3.6").release,
version.parse("0.3.7").release,
version.parse("0.3.8").release,
]:
# From: https://github.com/uqfoundation/dill/blob/dill-0.3.6/dill/_dill.py#L1104
@pklregister(CodeType)
def save_code(pickler, obj):
dill._dill.logger.trace(pickler, "Co: %s", obj)
############################################################################################################
# Modification here for huggingface/datasets
# The filename of a function is the .py file where it is defined.
# Filenames of functions created in notebooks or shells start with '<'
# ex: <ipython-input-13-9ed2afe61d25> for ipython, and <stdin> for shell
# Filenames of functions created in ipykernel the filename
# look like f"{tempdir}/ipykernel_{id1}/{id2}.py"
# Moreover lambda functions have a special name: '<lambda>'
# ex: (lambda x: x).__code__.co_name == "<lambda>" # True
#
# For the hashing mechanism we ignore where the function has been defined
# More specifically:
# - we ignore the filename of special functions (filename starts with '<')
# - we always ignore the line number
# - we only use the base name of the file instead of the whole path,
# to be robust in case a script is moved for example.
#
# Only those two lines are different from the original implementation:
co_filename = (
""
if obj.co_filename.startswith("<")
or (
len(obj.co_filename.split(os.path.sep)) > 1
and obj.co_filename.split(os.path.sep)[-2].startswith("ipykernel_")
)
or obj.co_name == "<lambda>"
else os.path.basename(obj.co_filename)
)
co_firstlineno = 1
# The rest is the same as in the original dill implementation, except for the replacements:
# - obj.co_filename => co_filename
# - obj.co_firstlineno => co_firstlineno
############################################################################################################
if hasattr(obj, "co_endlinetable"): # python 3.11a (20 args)
args = (
obj.co_lnotab, # for < python 3.10 [not counted in args]
obj.co_argcount,
obj.co_posonlyargcount,
obj.co_kwonlyargcount,
obj.co_nlocals,
obj.co_stacksize,
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_filename, # Modification for huggingface/datasets ############################################
obj.co_name,
obj.co_qualname,
co_firstlineno, # Modification for huggingface/datasets #########################################
obj.co_linetable,
obj.co_endlinetable,
obj.co_columntable,
obj.co_exceptiontable,
obj.co_freevars,
obj.co_cellvars,
)
elif hasattr(obj, "co_exceptiontable"): # python 3.11 (18 args)
args = (
obj.co_lnotab, # for < python 3.10 [not counted in args]
obj.co_argcount,
obj.co_posonlyargcount,
obj.co_kwonlyargcount,
obj.co_nlocals,
obj.co_stacksize,
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_filename, # Modification for huggingface/datasets ############################################
obj.co_name,
obj.co_qualname,
co_firstlineno, # Modification for huggingface/datasets #########################################
obj.co_linetable,
obj.co_exceptiontable,
obj.co_freevars,
obj.co_cellvars,
)
elif hasattr(obj, "co_linetable"): # python 3.10 (16 args)
args = (
obj.co_lnotab, # for < python 3.10 [not counted in args]
obj.co_argcount,
obj.co_posonlyargcount,
obj.co_kwonlyargcount,
obj.co_nlocals,
obj.co_stacksize,
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_filename, # Modification for huggingface/datasets ############################################
obj.co_name,
co_firstlineno, # Modification for huggingface/datasets #########################################
obj.co_linetable,
obj.co_freevars,
obj.co_cellvars,
)
elif hasattr(obj, "co_posonlyargcount"): # python 3.8 (16 args)
args = (
obj.co_argcount,
obj.co_posonlyargcount,
obj.co_kwonlyargcount,
obj.co_nlocals,
obj.co_stacksize,
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_filename, # Modification for huggingface/datasets ############################################
obj.co_name,
co_firstlineno, # Modification for huggingface/datasets #########################################
obj.co_lnotab,
obj.co_freevars,
obj.co_cellvars,
)
else: # python 3.7 (15 args)
args = (
obj.co_argcount,
obj.co_kwonlyargcount,
obj.co_nlocals,
obj.co_stacksize,
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_filename, # Modification for huggingface/datasets ############################################
obj.co_name,
co_firstlineno, # Modification for huggingface/datasets #########################################
obj.co_lnotab,
obj.co_freevars,
obj.co_cellvars,
)
pickler.save_reduce(dill._dill._create_code, args, obj=obj)
dill._dill.logger.trace(pickler, "# Co")
return
|