Jupyter dlopen cannot load any more object with static TLS

发表于 更新于 分类于 评论： 阅读次数：

A weird problem that execute "import torch" in bash works but when you run it in Jupyter notebook:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
ImportError Traceback (most recent call last)
<ipython-input-4-8ba1970b60ce> in <module>
 6 import random
 7 
----> 8  import torch
 9 import torch.nn as nn
 10 

~/miniconda3/envs/tf/lib/python3.6/site-packages/torch/__init__.py in <module>
 79 del _dl_flags
 80 
---> 81  from torch._C import *
 82 
 83 __all__ += [name for name in dir(_C)

ImportError: dlopen: cannot load any more object with static TLS

Seems that it's a compatibility issue. Even though there are a large volume of discussions but none of them works. Ironically, I accidently fixed the issue by resolving another: # RuntimeError cuDNN error CUDNN_STATUS_EXECUTION_FAILED Solution

Potential Solution

I marked these solutions as "potential" because I didn't make a strict control variable experiment. I tried all of them and it worked. :-)

My environment: 

1
2
3
4
5

1. Ubuntu 14.04 LTS
2. Python 3.6.9
3. PyTorch 1.3.0
4. cuda 10.1 (Tesla K40m)

The order of following solutions matters.

Upgrade/Downgrade Cuda

To the best of my understanding, this operation does the trick (but didn't verified). Downgrade cuda from 10.x to 9.0.

Upgrade/Downgrade PyTorch

1
conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch

If you use conda, you can refer to this page to find a working combination: INSTALLING PREVIOUS VERSIONS OF PYTORCH

Upgrade Jupyter

1
$ conda update jupyter

Upgrade Ubuntu

I do-release-upgrade from 14.04 to 16.04 and finally 18.04 LTS. Ubuntu Upgrades