Scikit-learn: ImportError: dlopen: cannot load any more object with static TLS with torch built with gcc 5.5

Created on 26 Jul 2019 · 12Comments · Source: scikit-learn/scikit-learn

I am not sure if this is a PyTorch bug, a scikit-learn bug or a numba, but this used to work in scikit-learn 0.20.3 and stopped working in the 0.21.0 series, so for now I am going to venture a guess that it is a regression in scikit learn.

When I do the following series of imports (minimized from the original import, which was import librosa), loading the following program fails:

import torch
import soundfile
import scipy.signal
import numba
import sklearn

with

Traceback (most recent call last):
  File "/opt/conda/lib/python3.6/site-packages/sklearn/__check_build/__init__.py", line 44, in <module>
    from ._check_build import check_build  # noqa
ImportError: dlopen: cannot load any more object with static TLS

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "test_torch.py", line 5, in <module>
    import sklearn
  File "/opt/conda/lib/python3.6/site-packages/sklearn/__init__.py", line 75, in <module>
    from . import __check_build
  File "/opt/conda/lib/python3.6/site-packages/sklearn/__check_build/__init__.py", line 46, in <module>
    raise_build_error(e)
  File "/opt/conda/lib/python3.6/site-packages/sklearn/__check_build/__init__.py", line 41, in raise_build_error
    %s""" % (e, local_dir, ''.join(dir_content).strip(), msg))
ImportError: dlopen: cannot load any more object with static TLS
___________________________________________________________________________
Contents of /opt/conda/lib/python3.6/site-packages/sklearn/__check_build:
_check_build.cpython-36m-x86_64-linux-gnu.so__pycache__               __init__.py
setup.py
___________________________________________________________________________
It seems that scikit-learn has not been built correctly.

If you have installed scikit-learn from source, please do not forget
to build the package before using it: run `python setup.py install` or
`make` in the source directory.

If you have used an installer, please check that it is suited for your
Python version, your operating system and your platform.

Downgrading to scikit-learn 0.20.3 makes the problem go away.

Versions

jenkins@260bf77532d0:~/workspace/test$ python
Python 3.6.8 |Anaconda, Inc.| (default, Dec 30 2018, 01:22:34) 
[GCC 7.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import sklearn; sklearn.show_versions()

System:
    python: 3.6.8 |Anaconda, Inc.| (default, Dec 30 2018, 01:22:34)  [GCC 7.3.0]
executable: /opt/conda/bin/python
   machine: Linux-4.15.0-29-generic-x86_64-with-debian-jessie-sid

BLAS:
    macros: SCIPY_MKL_H=None, HAVE_CBLAS=None
  lib_dirs: /opt/conda/lib
cblas_libs: mkl_rt, pthread

Python deps:
       pip: 19.1.1
setuptools: 41.0.1
   sklearn: 0.21.2
     numpy: 1.16.4
     scipy: 1.1.0
    Cython: None
    pandas: None

Also, you may be interested in:

jenkins@260bf77532d0:~/workspace/test$ pip list | grep numba
numba                  0.43.1         
jenkins@260bf77532d0:~/workspace/test$ pip list | grep torch
torch                  1.2.0a0+ab800ad

The build of torch must be done with gcc 5.5.0 to cause this problem; other versions of gcc are known not to cause this problem.

For ease of reproduction, you can use the following docker image ezyang/scikit-learn-tls-repro:1 https://cloud.docker.com/repository/registry-1.docker.io/ezyang/scikit-learn-tls-repro Once in, follow the reproduction instructions as described above. (EDIT At time of writing, the Docker image is still uploading. Should be done soon.)

Source

ezyang

👍6 🚀1 ❤1 🎉1 😄1

Most helpful comment

I solved it by import sklearn,then import tensorflow.The import order result in this error.

Ningshiqi on 25 May 2020

🚀4 👍3 🎉2

All 12 comments

Thanks for the report. How did you build/install sklearn?

amueller on 26 Jul 2019

pip install scikit-learn

ezyang on 26 Jul 2019

👎22

Do you have the log for that? Did it build from source or did you install a wheel?

amueller on 26 Jul 2019

Collecting scikit-learn                                                                           
  Using cached https://files.pythonhosted.org/packages/85/04/49633f490f726da6e454fddc8e938bbb5bfed
2001681118d3814c219b723/scikit_learn-0.21.2-cp36-cp36m-manylinux1_x86_64.whl

ezyang on 26 Jul 2019

👍1

@ezyang you may want to share the Dockerfile if that's possible.

If anyone is interested in reproducing this error the right docker incantation to use is something like this:

docker run -it ezyang/scikit-learn-tls-repro:1 bash

Note that you need to specify the tag i.e. 1 explicitly otherwise you get a cryptic error message (the 'latest' tag does not exist):

Unable to find image 'ezyang/scikit-learn-tls-repro:latest' locally
docker: Error response from daemon: manifest for ezyang/scikit-learn-tls-repro:latest not found.

I have no idea why this would happen, but I have seem numerous bug reports related to this e.g. with pytorch and OpenCV https://github.com/pytorch/pytorch/issues/2083 or OpenCV and Tensorflow https://github.com/tensorflow/models/issues/523. All in all I would guess that this is not a scikit-learn bug.

The fact that it depends on the order of import is fishy, for exemple this works in your docker image:

python -c 'import torch; import sklearn; import soundfile; import scipy.signal; import numba'

lesteve on 1 Aug 2019

Note I tried to reproduce inside a conda environment (inside your docker image for good measure) and could not (scikit-learn 0.21.2 and pytorch 1.1.0), so I guess this could be linked to some changes in pytorch dev version.

conda create -n test -c pytorch pytorch scikit-learn scipy numba scikit-learn -y
conda activate test
pip install soundfile
python -c 'import torch; import soundfile; import scipy.signal; import numba; import sklearn'

$ conda list
# packages in environment at /opt/conda/envs/test:
#
# Name                    Version                   Build  Channel
_libgcc_mutex             0.1                        main  
blas                      1.0                         mkl  
ca-certificates           2019.5.15                     0  
certifi                   2019.6.16                py37_1  
cffi                      1.12.3           py37h2e261b9_0  
cudatoolkit               10.0.130                      0  
intel-openmp              2019.4                      243  
joblib                    0.13.2                   py37_0  
libedit                   3.1.20181209         hc058e9b_0  
libffi                    3.2.1                hd88cf55_4  
libgcc-ng                 9.1.0                hdf63c60_0  
libgfortran-ng            7.3.0                hdf63c60_0  
libstdcxx-ng              9.1.0                hdf63c60_0  
llvmlite                  0.29.0           py37hd408876_0  
mkl                       2019.4                      243  
mkl-service               2.0.2            py37h7b6447c_0  
mkl_fft                   1.0.12           py37ha843d7b_0  
mkl_random                1.0.2            py37hd81dba3_0  
ncurses                   6.1                  he6710b0_1  
ninja                     1.9.0            py37hfd86e86_0  
numba                     0.45.0           py37h962f231_0  
numpy                     1.16.4           py37h7e9f1db_0  
numpy-base                1.16.4           py37hde5b4d6_0  
openssl                   1.1.1c               h7b6447c_1  
pip                       19.1.1                   py37_0  
pycparser                 2.19                     py37_0  
python                    3.7.3                h0371630_0  
pytorch                   1.1.0           py3.7_cuda10.0.130_cudnn7.5.1_0    pytorch
readline                  7.0                  h7b6447c_5  
scikit-learn              0.21.2           py37hd81dba3_0  
scipy                     1.3.0            py37h7c811a0_0  
setuptools                41.0.1                   py37_0  
six                       1.12.0                   py37_0  
soundfile                 0.10.2                   pypi_0    pypi
sqlite                    3.29.0               h7b6447c_0  
tk                        8.6.8                hbc83047_0  
wheel                     0.33.4                   py37_0  
xz                        5.2.4                h14c3975_4  
zlib                      1.2.11               h7b6447c_3

lesteve on 1 Aug 2019

I guess it would be useful and helpful to get a bisect on scikit-learn, if the problem reproduces on a dev build.

ezyang on 1 Aug 2019

Generally speaking, my feeling is that the expertise on this kind of problems is on the PyTorch side. Personally, I never heard about static TLS before and I would guess this is the case of many other core scikit-learn devs although I could be wrong about the last statement.

IIUC you originally saw the problem with scikit-learn 0.21.2 and a pytorch dev version. I can not reproduce the problem on scikit-learn 0.21.2 and pytorch 1.1.0 as noted in https://github.com/scikit-learn/scikit-learn/issues/14485#issuecomment-517195977. If I was to try to understand this in more details, I would bisect on PyTorch.

lesteve on 1 Aug 2019

The issue @ezyang linked has bunch of information on this TLS (thread local store) issue.
Here's some info I dug up before: https://github.com/pytorch/pytorch/issues/2575#issuecomment-369892859

;TLDR: Something in the chain of imports was not C/C++ compiled with -gPIC flag. Importing that library causes a problem that turns all imports to "static TLS". There is a maximum amount of such "static TLS" slots (names I use here are surely incorrect). Exact N of slots depends on the OS, and how it was compiled.

In the linked pytorch issue 2575, there is a mention that it is OpenMP which was compiled without the flag causing the cascade.
This scikit-learn issue might be due to some new library being introduced or some change, eating just few more static TLS slots.

Note: Not a real expert. There might be other sources for this error than "one/some lib missing `-gPIC' flag when it was compiled". Haven't found one though.