RuntimeError: Failed to import transformers.models.flaubert.modeling_flaubert because of the following error (look up to see its traceback): module 'signal' has no attribute 'SIGKILL'

[BenoitDalFerro]

System Info

• transformers version: 4.24.0
• Platform: Windows-10-10.0.19044-SP0
• Python version: 3.9.13
• Huggingface_hub version: 0.10.1
• PyTorch version (GPU?): 1.13.0 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?: Yes
• Using distributed or parallel set-up in script?: No

Who can help?

@sgugger, @D3xter1922
@malfet (we now come across each other daily, the world is so small...)

Opening here to document for the wider crowd, this is actually a PyTorch issue affecting Windows which went cf. pytorch/pytorch#85427

cf. Removed XLMModel inheritance from FlaubertModel(torch+tf)

cf in particular blame lines 26-45

error [caused by from ...modeling_utils import PreTrainedModel, SequenceSummary, SQuADHead line 37]
(https://github.com/huggingface/transformers/blob/main/src/transformers/models/flaubert/modeling_flaubert.py#L37)

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Opening here to document for the wider crowd, this is a PyTorch issue affecting Windows which went cf. pytorch/pytorch#85427

cf. complete stack trace for your reference only and closing

error [caused by from ...modeling_utils import PreTrainedModel, SequenceSummary, SQuADHead line 37]
(https://github.com/huggingface/transformers/blob/main/src/transformers/models/flaubert/modeling_flaubert.py#L37)

this goes down to the accelerate package

from transformers import (FlaubertWithLMHeadModel)

NOTE: Redirects are currently not supported in Windows or MacOs.
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
File ~\miniconda3\envs\MyEnv\lib\site-packages\transformers\utils\import_utils.py:1076, in _LazyModule._get_module(self, module_name)
   1075 try:
-> 1076     return importlib.import_module("." + module_name, self.__name__)
   1077 except Exception as e:

File ~\miniconda3\envs\MyEnv\lib\importlib\__init__.py:127, in import_module(name, package)    126         level += 1
--> 127 return _bootstrap._gcd_import(name[level:], package, level)

File <frozen importlib._bootstrap>:1030, in _gcd_import(name, package, level)

File <frozen importlib._bootstrap>:1007, in _find_and_load(name, import_)

File <frozen importlib._bootstrap>:986, in _find_and_load_unlocked(name, import_)

File <frozen importlib._bootstrap>:680, in _load_unlocked(spec)

File <frozen importlib._bootstrap_external>:850, in exec_module(self, module)

File <frozen importlib._bootstrap>:228, in _call_with_frames_removed(f, *args, **kwds)

File ~\miniconda3\envs\MyEnv\lib\site-packages\transformers\models\flaubert\modeling_flaubert.py:37, in <module>
     29 from ...modeling_outputs import (
     30     BaseModelOutput,
     31     MaskedLMOutput,
   (...)
     35     TokenClassifierOutput,
     36 )
---> 37 from ...modeling_utils import PreTrainedModel, SequenceSummary, SQuADHead
     38 from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer

File ~\miniconda3\envs\MyEnv\lib\site-packages\transformers\modeling_utils.py:78, in <module>
     77 if is_accelerate_available():
---> 78     from accelerate import __version__ as accelerate_version
     79     from accelerate import dispatch_model, infer_auto_device_map, init_empty_weights

File ~\miniconda3\envs\MyEnv\lib\site-packages\accelerate\__init__.py:7, in <module>
      5 __version__ = "0.13.2"
----> 7 from .accelerator import Accelerator
      8 from .big_modeling import cpu_offload, disk_offload, dispatch_model, init_empty_weights, load_checkpoint_and_dispatch

File ~\miniconda3\envs\MyEnv\lib\site-packages\accelerate\accelerator.py:27, in <module>
     25 import torch
---> 27 from .checkpointing import load_accelerator_state, load_custom_state, save_accelerator_state, save_custom_state
     28 from .data_loader import prepare_data_loader

File ~\miniconda3\envs\MyEnv\lib\site-packages\accelerate\checkpointing.py:24, in <module>
     22 from torch.cuda.amp import GradScaler
---> 24 from .utils import (
     25     MODEL_NAME,
     26     OPTIMIZER_NAME,
     27     RNG_STATE_NAME,
     28     SCALER_NAME,
     29     SCHEDULER_NAME,
     30     get_pretty_name,
     31     is_tpu_available,
     32     save,
     33 )
     36 if is_tpu_available(check_device=False):

File ~\miniconda3\envs\MyEnv\lib\site-packages\accelerate\utils\__init__.py:96, in <module>
     87     from .deepspeed import (
     88         DeepSpeedEngineWrapper,
     89         DeepSpeedOptimizerWrapper,
   (...)
     93         HfDeepSpeedConfig,
     94     )
---> 96 from .launch import PrepareForLaunch, _filter_args, get_launch_prefix
     97 from .memory import find_executable_batch_size

File ~\miniconda3\envs\MyEnv\lib\site-packages\accelerate\utils\launch.py:25, in <module>
     24 if is_torch_version(">=", "1.9.0"):
---> 25     import torch.distributed.run as distrib_run
     28 def get_launch_prefix():

File ~\miniconda3\envs\MyEnv\lib\site-packages\torch\distributed\run.py:386, in <module>
    385 from torch.distributed.elastic.utils.logging import get_logger
--> 386 from torch.distributed.launcher.api import LaunchConfig, elastic_launch
    389 log = get_logger()

File ~\miniconda3\envs\MyEnv\lib\site-packages\torch\distributed\launcher\__init__.py:10, in <module>
      1 #!/usr/bin/env/python3
      2 
      3 # Copyright (c) Facebook, Inc. and its affiliates.
   (...)
      6 # This source code is licensed under the BSD-style license found in the
      7 # LICENSE file in the root directory of this source tree.
---> 10 from torch.distributed.launcher.api import (  # noqa: F401
     11     LaunchConfig,
     12     elastic_launch,
     13     launch_agent,
     14 )

File ~\miniconda3\envs\MyEnv\lib\site-packages\torch\distributed\launcher\api.py:15, in <module>
     14 from torch.distributed.elastic import events, metrics
---> 15 from torch.distributed.elastic.agent.server.api import WorkerSpec
     16 from torch.distributed.elastic.agent.server.local_elastic_agent import LocalElasticAgent

File ~\miniconda3\envs\MyEnv\lib\site-packages\torch\distributed\elastic\agent\server\__init__.py:40, in <module>
     31 from .api import (  # noqa: F401
     32     ElasticAgent,
     33     RunResult,
   (...)
     38     WorkerState,
     39 )
---> 40 from .local_elastic_agent import TORCHELASTIC_ENABLE_FILE_TIMER, TORCHELASTIC_TIMER_FILE

File ~\miniconda3\envs\MyEnv\lib\site-packages\torch\distributed\elastic\agent\server\local_elastic_agent.py:19, in <module>
     17 from typing import Any, Dict, Optional, Tuple
---> 19 import torch.distributed.elastic.timer as timer
     20 from torch.distributed.elastic import events

File ~\miniconda3\envs\MyEnv\lib\site-packages\torch\distributed\elastic\timer\__init__.py:44, in <module>
     43 from .local_timer import LocalTimerClient, LocalTimerServer  # noqa: F401
---> 44 from .file_based_local_timer import FileTimerClient, FileTimerServer, FileTimerRequest

File ~\miniconda3\envs\MyEnv\lib\site-packages\torch\distributed\elastic\timer\file_based_local_timer.py:63, in <module>
     52         return json.dumps(
     53             {
     54                 "version": self.version,
   (...)
     59             },
     60         )
---> 63 class FileTimerClient(TimerClient):
     64     """
     65     Client side of ``FileTimerServer``. This client is meant to be used
     66     on the same host that the ``FileTimerServer`` is running on and uses
   (...)
     79                         negative or zero signal will not kill the process.
     80     """

File ~\miniconda3\envs\MyEnv\lib\site-packages\torch\distributed\elastic\timer\file_based_local_timer.py:81, in FileTimerClient()
     64 """
     65 Client side of ``FileTimerServer``. This client is meant to be used
     66 on the same host that the ``FileTimerServer`` is running on and uses
   (...)
     79                     negative or zero signal will not kill the process.
     80 """
---> 81 def __init__(self, file_path: str, signal=signal.SIGKILL) -> None:
     82     super().__init__()

AttributeError: module 'signal' has no attribute 'SIGKILL'

The above exception was the direct cause of the following exception:

RuntimeError                              Traceback (most recent call last)
Input In [3], in <cell line: 1>()
----> 1 from transformers import (PretrainedConfig, FlaubertConfig, AutoTokenizer, FlaubertTokenizer, FlaubertWithLMHeadModel, TrainingArguments, DataCollatorForLanguageModeling) #pipeline
      2 from datasets import (load_dataset, load_from_disk, concatenate_datasets, ClassLabel)
      3 import pytorch_lightning as pl

File <frozen importlib._bootstrap>:1055, in _handle_fromlist(module, fromlist, import_, recursive)

File ~\miniconda3\envs\MyEnv\lib\site-packages\transformers\utils\import_utils.py:1067, in _LazyModule.__getattr__(self, name)
   1065 elif name in self._class_to_module.keys():
   1066     module = self._get_module(self._class_to_module[name])
-> 1067     value = getattr(module, name)
   1068 else:
   1069     raise AttributeError(f"module {self.__name__} has no attribute {name}")

File ~\miniconda3\envs\MyEnv\lib\site-packages\transformers\utils\import_utils.py:1066, in _LazyModule.__getattr__(self, name)
   1064     value = self._get_module(name)
   1065 elif name in self._class_to_module.keys():
-> 1066     module = self._get_module(self._class_to_module[name])
   1067     value = getattr(module, name)
   1068 else:

File ~\miniconda3\envs\MyEnv\lib\site-packages\transformers\utils\import_utils.py:1078, in _LazyModule._get_module(self, module_name)
   1076     return importlib.import_module("." + module_name, self.__name__)
   1077 except Exception as e:
-> 1078     raise RuntimeError(
   1079         f"Failed to import {self.__name__}.{module_name} because of the following error (look up to see its"
   1080         f" traceback):\n{e}"
   1081     ) from e

RuntimeError: Failed to import transformers.models.flaubert.modeling_flaubert because of the following error (look up to see its traceback):
module 'signal' has no attribute 'SIGKILL'

Expected behavior

flawless import as usual

[BenoitDalFerro]

This is a PyTorch issue affecting Windows which went cf. pytorch/pytorch#85427

[BenoitDalFerro]

actually reopening as bug on Windows PyTorch side makes Transformers crash

[sgugger]

I'm not sure what you intend us to do to fix this, since it comes from PyTorch?

[BenoitDalFerro]

for pip would propose to add to requirement.txt torch<=1.12.1 ? and for conda feedstocks' environment.yaml pytorch<=1.12.1
point is Transformers do really crash on Windows with PyTorch=1.13.0

[sgugger]

PyTorch is already pinned in the setup.

[BenoitDalFerro]

indeed, not yet visible downstream (pip, conda) as of currentbut quite right https://github.com/huggingface/transformers/blame/main/setup.py#L166
#19989
closing