Unable to install torch_sparse==0.6.7 with correct CUDA version

[LucaMatarazzo]

Problem

Hi all, I am trying to install torch_sparse==0.6.7 in a virtual environment on a linux server to replicate the following paper. I can install torch_sparse==0.6.7 correctly in the venv but when trying to import it I get the following error:

Detected that PyTorch and torch_sparse were compiled with different CUDA versions. PyTorch has CUDA version 10.2 and torch_sparse has CUDA version 11.4. Please reinstall the torch_sparse that matches your PyTorch install.

Tried solutions

To solve it, I uninstalled it (pip uninstall torch_sparse) and tried the following solutions:

1. pip install torch-sparse==0.6.7+cu102 -f https://pytorch-geometric.com/whl/torch-1.6.0.html (as originally suggested here). But I get the following error:

ERROR: Could not find a version that satisfies the requirement torch-sparse==0.6.7+cu102 (from versions: latest+cpu, latest+cu92, latest+cu101, latest+cu102, 0.1.0, 0.2.0, 0.2.1, 0.2.2, 0.2.3, 0.2.4, 0.4.0, 0.4.1, 0.4.2, 0.4.3, 0.4.4, 0.5.1, 0.6.0, 0.6.1, 0.6.3, 0.6.4, 0.6.5, 0.6.6, 0.6.7, 0.6.7+cpu, 0.6.7+cu101, 0.6.7+cu102, 0.6.7+cu92, 0.6.8, 0.6.8+cpu, 0.6.8+cu101, 0.6.8+cu102, 0.6.8+cu92, 0.6.9, 0.6.10, 0.6.11, 0.6.12) ERROR: No matching distribution found for torch-sparse==0.6.7+cu102

2. pip install torch-scatter torch-sparse -f https://data.pyg.org/whl/torch-1.6.0+cu102.html (as indicated here). But I get the following error:

Looking in links: https://data.pyg.org/whl/torch-1.6.0+cu102.html Collecting torch-scatter Using cached torch_scatter-2.0.9.tar.gz (21 kB) Preparing metadata (setup.py) ... done Collecting torch-sparse Using cached torch_sparse-0.6.12.tar.gz (43 kB) Preparing metadata (setup.py) ... done Requirement already satisfied: scipy in ./.venv/lib/python3.6/site-packages (from torch-sparse) (1.5.4) Requirement already satisfied: numpy>=1.14.5 in ./.venv/lib/python3.6/site-packages (from scipy->torch-sparse) (1.19.2) Building wheels for collected packages: torch-scatter, torch-sparse Building wheel for torch-scatter (setup.py) ... error ERROR: Command errored out with exit status 1:
[...]
ERROR: Failed building wheel for torch-scatter Running setup.py clean for torch-scatter Building wheel for torch-sparse (setup.py) ... error ERROR: Command errored out with exit status 1:
[...]
ERROR: Failed building wheel for torch-sparse Running setup.py clean for torch-sparse Failed to build torch-scatter torch-sparse Installing collected packages: torch-sparse, torch-scatter Running setup.py install for torch-sparse ... error ERROR: Command errored out with exit status 1:
[...]
error: command 'gcc' failed with exit status 1 ---------------------------------------- ERROR: Command errored out with exit status 1: /home/myname/directoryname/.venv/bin/python3 -u -c 'import io, os, sys, setuptools, tokenize; sys.argv[0] = '"'"'/tmp/pip-install-uu1cqap8/torch-sparse_908976f98b4e496880675245c7dab3f0/setup.py'"'"'; __file__='"'"'/tmp/pip-install-uu1cqap8/torch-sparse_908976f98b4e496880675245c7dab3f0/setup.py'"'"';f = getattr(tokenize, '"'"'open'"'"', open)(__file__) if os.path.exists(__file__) else io.StringIO('"'"'from setuptools import setup; setup()'"'"');code = f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, __file__, '"'"'exec'"'"'))' install --record /tmp/pip-record-mwnswc97/install-record.txt --single-version-externally-managed --compile --install-headers /home/matarazzo/CFLP/.venv/include/site/python3.6/torch-sparse

Environment

I am using Python 3.6.8 (while the authors originally use 3.8.5), pip 21.3.1 and installed in the venv the following:

• torch_geometric==1.6.1
• torch==1.6.0
• matplotlib==3.3.1
• scikit_network==0.15.0
• numpy==1.19.2
• six==1.15.0
• networkx==2.5
• dgl==0.5.1
• scipy==1.5.4
• geomloss==0.2.4
• ogb==1.3.1
• scikit_learn==0.24.2
• wheel==0.37.1

[rusty1s]

Try

pip install torch-scatter torch-sparse --no-index -f https://data.pyg.org/whl/torch-1.6.0+cu102.html

[LucaMatarazzo]

Hi, thanks for the reply. My path to CUDA should be /usr/local/cuda. Inside, there is a lib64. LD_LIBRARY_PATH was empty before, and I defined as follows: export LD_LIBRARY_PATH=/usr/local/cuda/lib64:. Still, when running the script I run in the same error as before importing torch_sparse. I noticed that listing the contents of /usr/local/cuda/lib64, among others, I find inside the following:

libcusparse.so
libcusparse.so.11
libcusparse.so.11.6.0.120
libcusparse_static.a

However, there is no libcusparse.so.10. I'm not sure if that's the problem and how to solve it...

EDIT: also, running print(torch.version.cuda) I obtain 10.2, as expected, while running nvcc --version in the terminal I obtain:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Mon_Oct_11_21:27:02_PDT_2021
Cuda compilation tools, release 11.4, V11.4.152
Build cuda_11.4.r11.4/compiler.30521435_0

which suggests I have maybe multiple CUDA installations. Again, not sure on how to set things properly while keeping the correct versions of torch and torch_geometric (1.6.0 and 1.6.1). Note that I installed using pip install torch==1.6.0.

[rusty1s]

You can take a look in your installation folder of PyTorch, I would assume libcusparse.so.10 is included in there.

[LucaMatarazzo]

Unfortunately it seems that's not the case: with pip show torch I get:

Location: /home/myname/directoryname/.venv/lib/python3.6/site-packages

However, using find .venv -name "libcusparse*" in the terminal to look for the file inside the venv where torch was installed, I get nothing. I don't know if that's useful, but using find .venv -name "libc*", I get the following:

.venv/lib/python3.6/site-packages/torch/lib/libc10_cuda.so
.venv/lib/python3.6/site-packages/torch/lib/libcaffe2_detectron_ops_gpu.so
.venv/lib/python3.6/site-packages/torch/lib/libcaffe2_observers.so
.venv/lib/python3.6/site-packages/torch/lib/libc10.so
.venv/lib/python3.6/site-packages/torch/lib/libcaffe2_nvrtc.so
.venv/lib/python3.6/site-packages/torch/lib/libcudart-80664282.so.10.2
.venv/lib/python3.6/site-packages/torch/lib/libcaffe2_module_test_dynamic.so

[rusty1s]

Did you install torch via miniconda? You may also find it in miniconda/lib path.

[LucaMatarazzo]

EDIT: After some more trials, it worked. Thank you for the fast and useful support! For others with similar issues: I used find / -name "libcusparse.so.10" -print 2>/dev/null to search in the whole system for libcusparse.so.10 and found this path:

/usr/local/tmp/pisa/.conda/pkgs/cudatoolkit-10.2.89-hfd86e86_1/lib/libcusparse.so.10

updated LD_LIBRARY_PATH as follows: export LD_LIBRARY_PATH=/usr/local/tmp/pisa/.conda/pkgs/cudatoolkit-10.2.89-hfd86e86_1/lib: and then was able to import torch_sparse.

Unfortunately I ran in another issue with CUDA but that's for another discussion forum. Thanks again! :)