OSError: /lib/x86_64-linux-gnu/libm.so.6: version `GLIBC_2.27' not found (required by /home/user/anaconda3/envs/hbk_gcn/lib/python3.9/site-packages/torch_spline_conv/_basis_cuda.so)

[kkwd22]

📚 Installation

Environment

• OS:ubuntu 16.04
• Python version:3.9
• PyTorch version:1.10
• CUDA/cuDNN version:10.2
• GCC version:
• How did you try to install PyTorch Geometric and its extensions (wheel, source):
• Any other relevant information:

Checklist

• [x] I followed the installation guide.
• [ ] I cannot find my error message in the FAQ.
• [ ] I set up CUDA correctly and can compile CUDA code via nvcc.
• [ ] I do have multiple CUDA versions on my machine.

Additional context

[rusty1s]

How did you try to install the extension packages and what's your GCC version?

[hcw0098]

same error

• OS : ubuntu 16.04
• python : 3.8/3.9, try them both
• pytorch: 1.10.0
• CUDA 10.2
• GCC version 7.5.0
• PyG isntall : using "conda install pyg -c pyg -c conda-forge" import torch_geometric successfully, but when I try to import torch_geometric.nn， this error occurred

[rusty1s]

Does that help? https://github.com/NVIDIA/TensorRT/issues/1124

[hcw0098]

I solved it by degrading torch to 1.9.0, dont know if it helps

[Huni-ML]

I meet the same problem when using Pytorch 1.10 and pyg with cuda 11.3. I install with the following commands: conda create -n pasp_gnn pytorch torchvision torchaudio cudatoolkit=11.3 python=3 -c pytorch -c conda-forge -y conda activate pasp_gnn conda install pyg -c pyg -c conda-forge -y when I run from torch_geometric.nn.models.dimenet import SphericalBasisLayer, it gives the same error: OSError: /lib64/libm.so.6: version GLIBC_2.27' not found (required by .../lib/python3.9/site-packages/torch_spline_conv/_basis_cuda.so)`

But it will not happen if I install with Pytorch 1.9 Maybe a conflict problems between Pytorch 1.10 and torch_spline_conv?

[rusty1s]

What's your GLIBC version?

ldd --version

[Huni-ML]

ldd --version

ldd (GNU libc) 2.17

[rusty1s]

I think the main problem is that our wheels are compiled with Ubuntu 18.04 which comes with GLIBC 2.27 by default. Any chance you can upgrade to a higher version? Otherwise, you may need to install the extensions via manual compiling:

pip install torch-spline-conv

[Huni-ML]

I think the main problem is that our wheels are compiled with Ubuntu 18.04 which comes with GLIBC 2.27 by default. Any chance you can upgrade to a higher version? Otherwise, you may need to install the extensions via manual compiling:

pip install torch-spline-conv

Manual compiling pass. Thank you!

[chrisyeh96]

@rusty1s - Since the wheels for torch-spline-conv links against GLIBC, and some systems have older versions of GLIBC installed, would it be possible to have torch-spline-conv (pyg conda version) link against a conda glib package (e.g., from the anaconda channel or conda-forge channel)? That way, versions would be managed within conda instead of relying on the native GLIBC installation which sometimes cannot be updated for various reasons.

[rusty1s]

Yes, I think this is possible. Am I correct to assume that all we need to do is install glib via conda prior to building the package here?

[chrisyeh96]

Yes, I think this is possible. Am I correct to assume that all we need to do is install glib via conda prior to building the package here?

I wish I could help, but unfortunately I know very little about GitHub Actions and building packages. My original idea was that you could specify glib as a conda dependency of pyg.

[hnisonoff]

I am running into this issue and have been unable to resolve it by running conda install -c conda-forge glib before installing pytorch-spline-conv. Does anyone have suggestions for other solutions? Thanks!

[rusty1s]

For now, the only workaround is to install an older version of torch-spline-conv or avoid to install it in the first place (it's an optional dependency of PyG).

[yuchenlichuck]

I come across the same problems

[ljw-struggle]

I think the main problem is that our wheels are compiled with Ubuntu 18.04 which comes with GLIBC 2.27 by default. Any chance you can upgrade to a higher version? Otherwise, you may need to install the extensions via manual compiling:

pip install torch-spline-conv

How to install the extension via manual compiling?

[rusty1s]

pip install without the -f option will result in a manual compilation. It is similar to doing

git clone https://github.com/rusty1s/pytorch_spline_conv
cd pytorch_spline_conv
python setup.py install

[almumujin]

When I manually compiled torch-spline-conv：

r: invalid static_cast from type ‘const torch::OrderedDict<std::basic_string, std::shared_ptrtorch::nn::Module >’ to type ‘torch::OrderedDict<std::basic_string, std::shared_ptrtorch::nn::Module >&’ error: command '/usr/local/cuda-11.3/bin/nvcc' failed with exit code 1

[fengenxiang1123]

same error

OS : ubuntu 16.04 python : 3.9 pytorch: 1.10.2 CUDA 10.2 GCC version 5.4.0 PyG isntall : using "conda install pyg=2.0.4 -c pyg -c conda-forge" import torch_geometric successfully, but when I try to import torch_geometric.nn as pygnn， this error occurred

I solved this by "pin uninstall torch-spline-conv" and "pip install torch-spline-conv -f https://data.pyg.org/whl/torch-${TORCH}+${CUDA}.html"