problem loading the *_cpu.so file

[Mollyx06]

Hello, thanks for sharing your detailed tutorial,
I don't have a GPU environment configured on my computer, so I'm trying to install the CPU version of torch and related dependency packages as follows
/opt/software/python/bin/python3.8 -m pip install --upgrade pip
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-2.0.1+cpu.html
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-2.0.1+cpu.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-2.0.1+cpu.html
pip install torch-geometric==2.0

But when importing they
OSError: /opt/software/python/lib/python3.8/site-packages/torch_sparse/_sample_cpu.so: undefined symbol: _ZN2at4_ops11randint_low4callEllN3c108ArrayRefINS2_6SymIntEEENS2_8optionalINS2_10ScalarTypeEEENS6_INS2_6LayoutEEENS6_INS2_6DeviceEEENS6_IbEE

my pip version
torch 2.3.0+cpu
torch-cluster 1.6.3+pt20cpu
torch_geometric 2.0.0
torch-scatter 2.1.2+pt20cpu
torch-sparse 0.6.18+pt20cpu
torchaudio 2.3.0+cpu
torchvision 0.18.0+cpu

I'm wondering if there are any conflicts between these versions and how I can get it to recognise this cpu library?

[rusty1s]

This is likely because you are installing torch==2.3.0, but specify whl/torch-2.0.1+cpu.html in the links. You should install via

 pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-2.3.0+cpu.html

[Mollyx06]

Thank you for your reply! I changed the latest version and now the problem is solved!
pip install torch_geometric
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.3.0+cpu.html

[lamthuy]

I got a similar problem but the installation is for cuda when I created a fresh new conda env python 3.11 with the latest pytorch 2.3.0 with cuda 2.1.
Install PyG with

pip install torch_geometric
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.3.0+cu121.html

Then I got the following issue when I do

from torch_sparse import SparseTensor

376, in __init__
    self._handle = _dlopen(self._name, mode)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^
OSError: /lib64/libm.so.6: version `GLIBC_2.29' not found (required by anaconda3/envs/ok/lib/python3.11/site-packages/libpyg.so)

[lamthuy]

It turns out that the issue is due to pyglib, When install pyglib if I use:
pyg_lib==0.2 it will solve the problem, if I do not specify this old version, it will install 0.4 and the problem arises.