Ok, that's a weird one

Environment info

• transformers version: 4.1.1
• Platform: Linux-5.8.0-34-generic-x86_64-with-glibc2.32
• Python version: 3.9.0+
• PyTorch version (GPU?): 1.7.1 (False)
• Tensorflow version (GPU?): not installed (NA)
• Using GPU in script?: no
• Using distributed or parallel set-up in script?: no

Steps to reproduce the behavior:

1. Run the following script

   import torch
   import transformers
   
   loss = torch.tensor([1.0], requires_grad=True)
   loss.backward()

The script runs correctly but exits with

[1]    46823 segmentation fault (core dumped)  python testcase.py

Which doesn't happen if import transformers is commented out.

Only happens when on Python 3.9, it works as expected in 3.8.

Full env

certifi==2020.12.5
chardet==4.0.0
click==7.1.2
filelock==3.0.12
idna==2.10
joblib==1.0.0
numpy==1.19.4
packaging==20.8
pyparsing==2.4.7
regex==2020.11.13
requests==2.25.1
sacremoses==0.0.43
six==1.15.0
tokenizers==0.9.4
torch==1.7.1
tqdm==4.54.1
transformers==4.1.1
typing-extensions==3.7.4.3
urllib3==1.26.2