libcudart.so.10.0 ERROR

[usherbob]

I could install it successfully, but when I tried

from torch_geometric.data import Data,

there would be an error:

ImportError: libcudart.so.10.0: cannot open shared object file: No such file or directory

The complete error message is below:

>>> from torch_geometric.data import Data
 Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/bob/anaconda3/envs/torch_cu100/lib/python3.7/site-packages/torch_geometric/data/__init__.py", line 1, in <module>
    from .data import Data
  File "/home/bob/anaconda3/envs/torch_cu100/lib/python3.7/site-packages/torch_geometric/data/data.py", line 2, in <module>
    from torch_geometric.utils import (contains_isolated_nodes,
  File "/home/bob/anaconda3/envs/torch_cu100/lib/python3.7/site-packages/torch_geometric/utils/__init__.py", line 2, in <module>
    from .scatter import scatter_
  File "/home/bob/anaconda3/envs/torch_cu100/lib/python3.7/site-packages/torch_geometric/utils/scatter.py", line 1, in <module>
    import torch_scatter
  File "/home/bob/anaconda3/envs/torch_cu100/lib/python3.7/site-packages/torch_scatter/__init__.py", line 3, in <module>
    from .mul import scatter_mul
  File "/home/bob/anaconda3/envs/torch_cu100/lib/python3.7/site-packages/torch_scatter/mul.py", line 3, in <module>
    from torch_scatter.utils.ext import get_func
  File "/home/bob/anaconda3/envs/torch_cu100/lib/python3.7/site-packages/torch_scatter/utils/ext.py", line 5, in <module>
    import torch_scatter.scatter_cuda
ImportError: libcudart.so.10.0: cannot open shared object file: No such file or directory 

I installed CUDA-9.0 and CUDA-10.0 on my PC, and used a symlink pointing cuda package from /usr/local/cuda-10.0/ to /usr/local/cuda/. I wondered if it may cause the error.

[rusty1s]

Did you compile both PyTorch and its extensions with the same CUDA version? You can check this via torch.version.cuda

[usherbob]

Thanks for your reply. It may be the problem you mentioned above. I tested as you mentioned:

>>> torch.version.cuda
'9.0.176'

I guess there is something wrong with CUDA version.
However, I don't know how to compile Pytorch with its corresponding CUDA version. May you give me some help? Thanks.
I tried my way, delete the original symlink and point /usr/local/cuda-9.0 to /usr/local/cuda, but it did not work.

[RobinFrcd]

You're saying that /usr/local/cuda-9.0 points toward /usr/local/cuda, I think it should be the other way around.

[usherbob]

I mean I created a symlink with code,

sudo ln -s /usr/local/cuda-9.0/ /usr/local/cuda

This could be useful for me as I have multiple CUDA on my PC, I could switch CUDA version with this code.
Since I have installed PyTorch through conda install, I knew nothing about compiling PyTorch as @rusty1s said. How could I fix it?

[rusty1s]

PyTorch was compiled with CUDA version 9. As far as I know, you can change that by reinstalling PyTorch from conda and selecting the right CUDA version. For installing PyG, your symlinked version must match the version of PyTorch (in your case 9). However, you need to recompile the extension packages (which get simply cached when you try a reinstall) with the right CUDA version, just changing the symlink is not enough.

[usherbob]

Many thanks. I'll have a try.

[rusty1s]

I created a FAQ for common installation errors. Let me know if this fixes your issues.

[RuihongQiu]

@rusty1s I have the similar problem with PyTorch compiled with CUDA 10. And my environment is using CUDA 10 also.

from torch_geometric.data import Data
4 if torch.cuda.is_available():                                             
----> 5     import torch_scatter.scatter_cuda                                      
6                                                                            
7                                                                                                                                                     
ImportError: libcudart.so.9.0: cannot open shared object file: No such file or directory                                                                              
In [3]: torch.version.cuda                                                         
Out[3]: '10.0.130'

[rusty1s]

Did you modify LD_LIBRARY_PATH to point to CUDA 10?

[RuihongQiu]

I have done it and my PyTorch is working all well.

[rusty1s]

Maybe this helps?

[RuihongQiu]

It doesn't work.

The most strange thing is that I use all packages with CUDA 10.0 but here it still needs to look for libcudart.so.9.0. The version doesn't match. It should not be like this.

@rusty1s

[RuihongQiu]

Fix it.

torch-scatter seems to constantly find *.so files from

/usr/local/cuda/lib64

And there are CUDA 9.0 10.0 and 10.1 on my server. /usr/local/cuda/lib64 is originally set to link CUDA 9.0. Though my ~/.bashrc has set to link /usr/local/cuda-10.0/lib64 for LD_LIBRARY_PATH. The installation still finds /usr/local/cuda/lib64.

So I just remove /usr/local/cuda, copy /usr/local/cuda-10.0 and rename it as /usr/local/cuda, without changing anything in ~/.bashrc. Uninstall and reinstall. Fix it.

And then torch-sparse will cause another libcusparse.so.10.0 problem. So I change lines in ~/.bashrc to point to /usr/local/cuda/. And all things done.

[rusty1s]

Really strange error, but glad that it‘s now working.

[usherbob]

Sorry for my late reply. I still believe these errors

ImportError: libcudart.so.xxx: cannot open shared object file: No such file or directory

arise from multiple Cuda version. When I opened that issue, I have tried to change Cuda from 9.0 to 10.0. However, I might have forgotten to source ~/.bashrc , which could be the reason that changing Cuda version did not work. Anyway, I did not reinstall pytorch or pytorch_geometric package, it just works now.

[silent567]

Hi, problem fixed. I have checked the setup.py and found the cuda_path is collected by pytorch.
pytorch checked three paths sequentially: CUDA_HOME, /usr/local/cuda, and at last nvcc's path.

So, a quick work-around is to set the environment variable CUDA_HOME by

export CUDA_HOME=<path>

The related issue: pytorch/pytorch#22844

done

[FreyWang]

Did you modify LD_LIBRARY_PATH to point to CUDA 10?

really thank your solution!! Oh my god I was totally go crazy

[Silva-asilla]

FAQ

File not found :(