"Error rendering output item using 'jupyter-ipywidget-renderer'" when running transformers Trainer.train() on Python 3.13

[guangdawang]

Description

When running Hugging Face transformers' Trainer.train() in Jupyter Notebook, the progress bar fails to render with the error:

Error rendering output item using 'jupyter-ipywidget-renderer'
Cannot read properties of undefined (reading 'ipywidgetsKernel')

The training continues in the background (CPU usage is normal), but the progress bar and loss values are not visible. Adding os.environ['JUPYTER_DISABLE_WIDGETS'] = '1' works around the issue (but loses progress bar functionality).

Reproduce

1. Create a fresh environment:

   conda create -n ipywidgets-test python=3.13
   conda activate ipywidgets-test
   pip install jupyter ipywidgets transformers torch

2. Start Jupyter Notebook:

   jupyter notebook

3. Create a new notebook and run:

   from transformers import Trainer, TrainingArguments, GPT2LMHeadModel, GPT2Tokenizer
   
   # Load model
   model = GPT2LMHeadModel.from_pretrained('gpt2')
   tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
   tokenizer.pad_token = tokenizer.eos_token
   
   # Create training arguments
   training_args = TrainingArguments(
       output_dir='./test',
       num_train_epochs=1,
       per_device_train_batch_size=1,
       logging_steps=1,
   )
   
   # Create trainer (progress bar should appear even without dataset)
   trainer = Trainer(
       model=model,
       args=training_args,
   )
   
   # Run training - error appears here
   trainer.train()

4. Observe the red error message in the progress bar area

Expected behavior

The progress bar should display normally, showing:

• Epoch progress (e.g., 1/3)
• Current step
• Loss values
• Estimated time remaining

Context

• ipywidgets version: 8.1.8
• Operating System: WSL2 (Linux 6.6.87.2-microsoft-standard-WSL2) on Windows 11
• Browser: Chrome latest (also tested on Firefox, same issue)

Troubleshoot Output

Selected Jupyter core packages...
IPython          : 9.11.0
ipykernel        : 7.2.0
ipywidgets       : 8.1.8
jupyter_client   : 8.8.0
jupyter_core     : 5.9.1
jupyter_server   : 2.17.0
jupyterlab       : 4.5.6
nbclient         : 0.10.4
nbconvert        : 7.17.0
nbformat         : 5.10.4
notebook         : 7.5.5
traitlets        : 5.14.3
Other key package versions:

transformers: 5.3.0
torch: 2.10.0
Python: 3.13.12

Command Line Output

Start command: jupyter notebook
Output: [I 2026-03-13 10:30:00.123 LabApp] JupyterLab 4.5.6 is running at:
[I 2026-03-13 10:30:00.123 LabApp] http://localhost:8888/lab
[I 2026-03-13 10:30:00.123 LabApp] or http://127.0.0.1:8888/lab
(Normal startup, no errors)

Browser Output

Browser console error:
Error rendering output item using 'jupyter-ipywidget-renderer'
Cannot read properties of undefined (reading 'ipywidgetsKernel')
    at w (https://localhost:8888/static/lab/vendors-node_modules_@jupyter-widgets_html-manager_lib_index_js.xxx.js:1:xxxxx)
    ...

If using JupyterLab

• JupyterLab version: 4.5.6

Installed Labextensions

JupyterLab v4.5.6
(no additional labextensions, pure default installation)

Additional Notes

1. Issue only occurs with training progress bar: Other widgets in the notebook (like simple IPython.display) work fine
2. Temporary workaround works: Setting os.environ['JUPYTER_DISABLE_WIDGETS'] = '1' disables widgets and training proceeds normally (but with text output instead of progress bar)
3. Possible relevant factors:
   • Python 3.13 is very new (released in 2025)
   • transformers 5.3.0 is also the latest version
   • ipywidgets 8.x + Jupyter 7.x combination might have compatibility issues with new Python versions
4. Suggestions:
   • Could ipywidgets be tested with Python 3.13?
   • Could transformers provide a fallback option to force simple text progress bar?