Import error of the newest version of transformers

[smartliuhw]

System Info

system: Linux
transformers version: 4.39.1
torch version: 1.13.1
cuda version:1.16

Who can help?

@muellerzr and @pacman100

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

1. use the lm-evaluation-harness to run evaluation tasks
2. run Llama2-7b-hf on any of those tasks: nq_open, trivalqa, truthfulqa
3. get those results:

Traceback (most recent call last):                                                                      
  File "/usr/local/lib/python3.8/site-packages/transformers/utils/import_utils.py", line 1472, in _get_m
odule                                                                                                   
    return importlib.import_module("." + module_name, self.__name__)                                    
  File "/usr/lib64/python3.8/importlib/__init__.py", line 127, in import_module                         
    return _bootstrap._gcd_import(name[level:], package, level)                                         
  File "<frozen importlib._bootstrap>", line 1014, in _gcd_import                                       
  File "<frozen importlib._bootstrap>", line 991, in _find_and_load                                     
  File "<frozen importlib._bootstrap>", line 975, in _find_and_load_unlocked                            
  File "<frozen importlib._bootstrap>", line 671, in _load_unlocked                                     
  File "<frozen importlib._bootstrap_external>", line 783, in exec_module                               
  File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed                          
  File "/usr/local/lib/python3.8/site-packages/transformers/pipelines/__init__.py", line 47, in <module>
    from .audio_classification import AudioClassificationPipeline                                       
  File "/usr/local/lib/python3.8/site-packages/transformers/pipelines/audio_classification.py", line 21,
 in <module>                                                                                            
    from .base import Pipeline, build_pipeline_init_args                                                
  File "/usr/local/lib/python3.8/site-packages/transformers/pipelines/base.py", line 34, in <module>    
    from ..modelcard import ModelCard                                                                   
  File "/usr/local/lib/python3.8/site-packages/transformers/modelcard.py", line 48, in <module>         
    from .training_args import ParallelMode                                                             
  File "/usr/local/lib/python3.8/site-packages/transformers/training_args.py", line 75, in <module>     
    from .trainer_pt_utils import AcceleratorConfig                                                     
  File "/usr/local/lib/python3.8/site-packages/transformers/trainer_pt_utils.py", line 37, in <module>  
    from torch.optim.lr_scheduler import LRScheduler                                                    
ImportError: cannot import name 'LRScheduler' from 'torch.optim.lr_scheduler' (/usr/local/lib64/python3.
8/site-packages/torch/optim/lr_scheduler.py)   

Expected behavior

just finish the task

[smartliuhw]

System Info

system: Linux transformers version: 4.39.1 torch version: 1.13.1 cuda version:1.16

Who can help?

@muellerzr and @pacman100

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

1. use the lm-evaluation-harness to run evaluation tasks
2. run Llama2-7b-hf on any of those tasks: nq_open, trivalqa, truthfulqa
3. get those results:

Traceback (most recent call last):                                                                      
  File "/usr/local/lib/python3.8/site-packages/transformers/utils/import_utils.py", line 1472, in _get_m
odule                                                                                                   
    return importlib.import_module("." + module_name, self.__name__)                                    
  File "/usr/lib64/python3.8/importlib/__init__.py", line 127, in import_module                         
    return _bootstrap._gcd_import(name[level:], package, level)                                         
  File "<frozen importlib._bootstrap>", line 1014, in _gcd_import                                       
  File "<frozen importlib._bootstrap>", line 991, in _find_and_load                                     
  File "<frozen importlib._bootstrap>", line 975, in _find_and_load_unlocked                            
  File "<frozen importlib._bootstrap>", line 671, in _load_unlocked                                     
  File "<frozen importlib._bootstrap_external>", line 783, in exec_module                               
  File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed                          
  File "/usr/local/lib/python3.8/site-packages/transformers/pipelines/__init__.py", line 47, in <module>
    from .audio_classification import AudioClassificationPipeline                                       
  File "/usr/local/lib/python3.8/site-packages/transformers/pipelines/audio_classification.py", line 21,
 in <module>                                                                                            
    from .base import Pipeline, build_pipeline_init_args                                                
  File "/usr/local/lib/python3.8/site-packages/transformers/pipelines/base.py", line 34, in <module>    
    from ..modelcard import ModelCard                                                                   
  File "/usr/local/lib/python3.8/site-packages/transformers/modelcard.py", line 48, in <module>         
    from .training_args import ParallelMode                                                             
  File "/usr/local/lib/python3.8/site-packages/transformers/training_args.py", line 75, in <module>     
    from .trainer_pt_utils import AcceleratorConfig                                                     
  File "/usr/local/lib/python3.8/site-packages/transformers/trainer_pt_utils.py", line 37, in <module>  
    from torch.optim.lr_scheduler import LRScheduler                                                    
ImportError: cannot import name 'LRScheduler' from 'torch.optim.lr_scheduler' (/usr/local/lib64/python3.
8/site-packages/torch/optim/lr_scheduler.py)   

Expected behavior

just finish the task

after i downgrade transformers to version 4.38.2, the code works well. maybe there's a bug in the newest version

[Phynon]

Looks like the latest release (>= 4.39) breaks compatibility with pytorch 1.13.1 because torch/optim/lr_scheduler.py in that version does not export the name LRScheduler but transformers/trainer_pt_utils.py explicitly imports it.

transformers/src/transformers/trainer_pt_utils.py

Line 37 in 39114c0

from torch.optim.lr_scheduler import LRScheduler

[smartliuhw]

Looks like the latest release (>= 4.39) breaks compatibility with pytorch 1.13.1 because torch/optim/lr_scheduler.py in that version does not export the name LRScheduler but transformers/trainer_pt_utils.py explicitly imports it.

transformers/src/transformers/trainer_pt_utils.py

Line 37 in 39114c0

from torch.optim.lr_scheduler import LRScheduler

You are right. Since my cuda version is 11.6 i just can't find out if the higher pytorch version existing the same problem😂

[ArthurZucker]

That's a regression, @younesbelkada #29588 broke this 😢 sorry.
Do you want to open a fix for this? Protecting the import with some torch version available?

[smartliuhw]

That's a regression, @younesbelkada #29588 broke this 😢 sorry. Do you want to open a fix for this? Protecting the import with some torch version available?

yes, that would be great

[amyeroberts]

Ciosing as resolved in #29919.

Working from a source install should work in the meantime whilst we prepare a patch release: pip install git+https://github.com/huggingface/transformers

[xlinsz]

change LRScheduler to _LRScheduler, the function is defined as a internal function

[rphes]

This issue seems to be back again in 4.48.0. Our pipeline started failing today and downgrading to 4.47.1 resolved the issue.

[Rocketknight1]

Hi @rphes, this may be caused by your version of torch being older than 2.0. We have almost fully deprecated support for PyTorch 1.x at this point, and if you're on an older version we strongly recommend updating!

Can you confirm your torch version? If you're getting these issues with torch >= 2.0 then we'll investigate, but if not then we're unlikely to fix this.

[rphes]

Appreciate the reply @Rocketknight1. Indeed, like the original submitter, I'm on torch 1.13.1. I can see why you would deprecate torch 1.x, but I'm surprised it has seemingly stopped working after a minor version upgrade. In any case, the issue is easily resolved by pinning the version of transformers, but I just wanted to let you know I ran into this.