torch-sparse can't install with pytorch1.1.0+cuda9.2

LingxiaoShawn commented 5 years ago

📚 Installation

Environment

OS: Ubuntu 18.04
Python version: 3.6
PyTorch version: 1.0
CUDA/cuDNN version: 9.2
GCC version: gcc (Ubuntu 7.3.0-27ubuntu1~18.04) 7.3.0
How you tried to install PyTorch Geometric and its extensions (pip, source): I successfully installed pytorch-geometric and all extension with pytorch0.4+cuda9.2. But I want to use pytorch1.0 with cuda9.2. So I first intall pytorch1.1.0 with cuda9.2 from github source, since originally pytorch1.0 can only be installed with cuda9.0. Then: $ pip install --verbose --no-cache-dir torch-scatter [succeeded] $ pip install --verbose --no-cache-dir torch-sparse [failed]
Any other relevant information:

Checklist

[Yes] I followed the installation guide.
[Yes] I cannot find my error message in the FAQ.
[Yes] I set up CUDA correctly and can compile CUDA code via nvcc.
[Yes] I have cloned the repository and tried a manual installation from source.
[No] I do have multiple CUDA versions on my machine.

Additional context

pip install --no-cache-dir torch-sparse Collecting torch-sparse Downloading https://files.pythonhosted.org/packages/73/72/e374662f6f47d9ac0e082a6d5c18d14e15c52863e89c6bc6957a0d2ed026/torch_sparse-0.2.4.tar.gz Requirement already satisfied: scipy in ./anaconda3/envs/torch1.0/lib/python3.6/site-packages (from torch-sparse) (1.2.1) Requirement already satisfied: numpy>=1.8.2 in ./anaconda3/envs/torch1.0/lib/python3.6/site-packages (from scipy->torch-sparse) (1.16.2) Installing collected packages: torch-sparse Running setup.py install for torch-sparse ... error Complete output from command /home/lingxiao/anaconda3/envs/torch1.0/bin/python -u -c "import setuptools, tokenize;file='/tmp/pip-install-_oj5tmie/torch-sparse/setup.py';f=getattr(tokenize, 'open', open)(file);code=f.read().replace('\r\n', '\n');f.close();exec(compile(code, file, 'exec'))" install --record /tmp/pip-record-kvljuamw/install-record.txt --single-version-externally-managed --compile: running install running build running build_py creating build creating build/lib.linux-x86_64-3.6 creating build/lib.linux-x86_64-3.6/test copying test/test_coalesce.py -> build/lib.linux-x86_64-3.6/test copying test/utils.py -> build/lib.linux-x86_64-3.6/test copying test/test_transpose.py -> build/lib.linux-x86_64-3.6/test copying test/test_eye.py -> build/lib.linux-x86_64-3.6/test copying test/test_spspmm.py -> build/lib.linux-x86_64-3.6/test copying test/test_spmm.py -> build/lib.linux-x86_64-3.6/test copying test/init.py -> build/lib.linux-x86_64-3.6/test creating build/lib.linux-x86_64-3.6/torch_sparse copying torch_sparse/spmm.py -> build/lib.linux-x86_64-3.6/torch_sparse copying torch_sparse/transpose.py -> build/lib.linux-x86_64-3.6/torch_sparse copying torch_sparse/coalesce.py -> build/lib.linux-x86_64-3.6/torch_sparse copying torch_sparse/eye.py -> build/lib.linux-x86_64-3.6/torch_sparse copying torch_sparse/spspmm.py -> build/lib.linux-x86_64-3.6/torch_sparse copying torch_sparse/init.py -> build/lib.linux-x86_64-3.6/torch_sparse creating build/lib.linux-x86_64-3.6/torch_sparse/utils copying torch_sparse/utils/unique.py -> build/lib.linux-x86_64-3.6/torch_sparse/utils copying torch_sparse/utils/init.py -> build/lib.linux-x86_64-3.6/torch_sparse/utils running build_ext building 'torch_sparse.spspmm_cuda' extension creating build/temp.linux-x86_64-3.6 creating build/temp.linux-x86_64-3.6/cuda gcc -pthread -B /home/lingxiao/anaconda3/envs/torch1.0/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/home/lingxiao/anaconda3/envs/torch1.0/lib/python3.6/site-packages/torch/include -I/home/lingxiao/anaconda3/envs/torch1.0/lib/python3.6/site-packages/torch/include/torch/csrc/api/include -I/home/lingxiao/anaconda3/envs/torch1.0/lib/python3.6/site-packages/torch/include/TH -I/home/lingxiao/anaconda3/envs/torch1.0/lib/python3.6/site-packages/torch/include/THC -I/usr/local/cuda/include -I/home/lingxiao/anaconda3/envs/torch1.0/include/python3.6m -c cuda/spspmm.cpp -o build/temp.linux-x86_64-3.6/cuda/spspmm.o -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=spspmm_cuda -D_GLIBCXX_USE_CXX11_ABI=1 -std=c++11 cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++ /usr/local/cuda/bin/nvcc -I/home/lingxiao/anaconda3/envs/torch1.0/lib/python3.6/site-packages/torch/include -I/home/lingxiao/anaconda3/envs/torch1.0/lib/python3.6/site-packages/torch/include/torch/csrc/api/include -I/home/lingxiao/anaconda3/envs/torch1.0/lib/python3.6/site-packages/torch/include/TH -I/home/lingxiao/anaconda3/envs/torch1.0/lib/python3.6/site-packages/torch/include/THC -I/usr/local/cuda/include -I/home/lingxiao/anaconda3/envs/torch1.0/include/python3.6m -c cuda/spspmm_kernel.cu -o build/temp.linux-x86_64-3.6/cuda/spspmm_kernel.o -DCUDA_NO_HALF_OPERATORS -DCUDA_NO_HALF_CONVERSIONS -D__CUDA_NO_HALF2_OPERATORS__ --compiler-options '-fPIC' -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=spspmm_cuda -D_GLIBCXX_USE_CXX11_ABI=1 -std=c++11 cuda/spspmm_kernel.cu(78): error: no suitable user-defined conversion from "at::DeprecatedTypeProperties" to "const at::Type" exists

1 error detected in the compilation of "/tmp/tmpxft_00002ac8_00000000-6_spspmm_kernel.cpp1.ii".
error: command '/usr/local/cuda/bin/nvcc' failed with exit status 1

----------------------------------------

Command "/anaconda3/envs/torch1.0/bin/python -u -c "import setuptools, tokenize;file='/tmp/pip-install-_oj5tmie/torch-sparse/setup.py';f=getattr(tokenize, 'open', open)(file);code=f.read().replace('\r\n', '\n');f.close();exec(compile(code, file, 'exec'))" install --record /tmp/pip-record-kvljuamw/install-record.txt --single-version-externally-managed --compile" failed with error code 1 in /tmp/pip-install-_oj5tmie/torch-sparse/

rusty1s commented 5 years ago

Thank you for your detailed installation report. This is a very strange error and may be caused by internal ATen changes in PyTorch from v1.0 to v1.1. I generally try my best to keep all packages be installable with the stable PyTorch release, but can not guarantee that everything works smoothly in current PyTorch master. I suggest you to simple skip the compilation of spspmm_kernel.cu (it is not needed for PyG):

Open setup.py in torch-sparse and remove the line 16-18.
Open torch-sparse/__init__.py and remove line 6 and 20.
Install the package by running: python setup.py install

LingxiaoShawn commented 5 years ago

Hi Matthias it works without spspmm, thank you for help!

pyg-team / pytorch_geometric