Undefined symbol pytorch 1.7.0 and above

1

I was working with PyTorch 1.5.0 for quite some time and decided to update to 1.9.0.

System: Ubuntu 18.04
Kernel: 4.15.0-147-generic #151-Ubuntu SMP
I use pyenv 2.0.3 to manage virtual environment and python versions.
I use python 3.7.9 installed with pyenv.

Make a new virtual environment with pyenv called XXXX.

My pip env shows

$ pip --version
pip 20.1.1 from /home/clement/.pyenv/versions/3.7.9/envs/XXXX/lib/python3.7/site-packages/pip (python 3.7)
$ pip uninstall torch torchvision torchaudio
WARNING: Skipping torch as it is not installed.
WARNING: Skipping torchvision as it is not installed.
WARNING: Skipping torchaudio as it is not installed.
$ python3 -c "import torch; torch.__version__"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
ModuleNotFoundError: No module named 'torch'

Everything is clean with the new setup.
Now install PyTorch 1.9.0 everything goes well no error message from pip

pip3 install torch==1.9.0+cpu torchvision==0.10.0+cpu torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html

So everything looks fine but

$ python3 -c "import torch; torch.__version__"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/home/clement/.pyenv/versions/3.7.9/envs/XXXX/lib/python3.7/site-packages/torch/__init__.py", line 197, in <module>
    from torch._C import *  # noqa: F403
ImportError: /home/clement/.pyenv/versions/3.7.9/envs/XXXX/lib/python3.7/site-packages/torch/lib/libtorch_cpu.so: undefined symbol: _ZN3c1019UndefinedTensorImpl10_singletonE

I found that any version above 1.7.0 (included) give me this kind of error message.

  • 1.7.0+cpu: _ZTIN3c1010TensorImplE
  • 1.7.1+cpu: _ZTIN3c1010TensorImplE
  • 1.8.0+cpu: _ZTIN3c1010TensorImplE
  • 1.9.0+cpu: _ZN3c1019UndefinedTensorImpl10_singletonE

I also tried the “normal” version pip install torch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0

  • 1.7.0: _ZN3c107Warning4warnENS_14SourceLocationERKSsb
    I don’t have any compatible GPU so maybe the above error is as expected?

I also tried with the python 3.6.9 from the ubuntu 18.04 repository I have the same behavior as above.

Furthermore, I tried to clean anything on my system that contains find / -iname *torch*.

Honestly I don’t know where to look for this issue. Any version prior to 1.7.0 install and works without any issues.
I use docker with Ubuntu 18.04, and I could install 1.7.1 and works with torch. I think this come from my workstation, but I have no clues. If someone knows where I can start to look at.

1 Like
2

Usually you would get this error message if you are trying to import custom extensions before importing PyTorch, which doesn’t seem to be the case here.
I also assume that you don’t have any files, which could cause conflicts, in the current working directory.
In that case, could you create a new virtual environment, install PyTorch there, and test it again?

3

No directory/file was named torch on my working folder, if it was the case I assume that torch==1.6.0 will also fail no ?

Here are the steps I have done for the new venv:

$ mkdir new_venv
$ cd new_venv
$ virtualenv -p python3 pytorch_test
Running virtualenv with interpreter /home/clement/.pyenv/shims/python3
Using base prefix '/usr'
New python executable in /home/clement/new_env/pytorch_test/bin/python3
Also creating executable in /home/clement/new_env/pytorch_test/bin/python
Installing setuptools, pkg_resources, pip, wheel...done.
$ source pytorch_test/bin/activate
(pytorch_test) $ where python3
/home/clement/new_env/pytorch_test/bin/python3
/usr/bin/python3
(pytorch_test) $ pip freeze
pkg-resources==0.0.0
(pytorch_test) $ pip install torch==1.7.0+cpu torchvision==0.8.0 torchaudio==0.7.0 -f https://download.pytorch.org/whl/torch_stable.html
(pytorch_test) $ python3 -c "import torch; print(torch.__version__)"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/home/clement/new_env/pytorch_test/lib/python3.6/site-packages/torch/__init__.py", line 190, in <module>
    from torch._C import *
ImportError: /home/clement/new_env/pytorch_test/lib/python3.6/site-packages/torch/lib/libtorch_cpu.so: undefined symbol: _ZTIN3c1010TensorImplE
(pytorch_test) $ pip uninstall torch torchvision torchaudio
(pytorch_test) $ 
(pytorch_test) $ pip install torch==1.6.0+cpu torchvision==0.7.0+cpu -f https://download.pytorch.org/whl/torch_stable.html
(pytorch_test) $ python3 -c "import torch; print(torch.__version__)"
1.6.0+cpu
4

The only way I found to make it work is to use a docker container with the required PyTorch version in it.

1 Like