passionlab / cagnet Goto Github PK

Hi,
I successfully built and tested CAGNET on the single node.

Would it be possible to run CAGNET without ddlrun on the distributed system?

I am trying to run this on the TACC, Frontera server (link), but it does not have any IBM package, and it seems like installing them requires sudo permission.

I am not familiar with ddlrun. If I understand this correctly, it constructs efficient/effective mpirun command by considering the target environment, don't it? Then, could you please provide any corresponding mpirun to ddlrun?
For example, I would like to run CAGNET, reddit training on the 2 servers with 8 GPUs (4 GPUs per each server)

I have created a similar environment to reproduce the results in the CAGNET paper.
My environment is:
torch-1.4.0 torch-cluster-1.4.5 torch-geometric-1.3.2 torch-scatter-1.4.0 torch-sparse-0.4.3 torchvision-0.5.0 CUDA10.1 python-3.6.10
You can notice that only the version of pytorch is different from yours. I succeeded in building sparse-extension. Then I run python gcn_distr_15d.py --graphname=Reddit --download=True. It reports :
Traceback (most recent call last): File "gcn_distr_15d.py", line 27, in <module> from sparse_coo_tensor_cpp import sparse_coo_tensor_gpu, spmm_gpu ImportError: /root/miniconda3/envs/cagnet/lib/python3.6/site-packages/sparse_coo_tensor_cpp-0.0.0-py3.6-linux-x86_64.egg/sparse_coo_tensor_cpp.cpython-36m-x86_64-linux-gnu.so: undefined symbol: cusparseSpMM_bufferSize

It's quite a strange error. It would be nice if you can give some suggestions to solve my problem.

Hello,

I got this error when compiling it:

$ python setup.py install
running install
running bdist_egg
running egg_info
writing sparse_coo_tensor_cpp.egg-info/PKG-INFO
writing dependency_links to sparse_coo_tensor_cpp.egg-info/dependency_links.txt
writing top-level names to sparse_coo_tensor_cpp.egg-info/top_level.txt
reading manifest file 'sparse_coo_tensor_cpp.egg-info/SOURCES.txt'
writing manifest file 'sparse_coo_tensor_cpp.egg-info/SOURCES.txt'
installing library code to build/bdist.linux-x86_64/egg
running install_lib
running build_ext
building 'sparse_coo_tensor_cpp' extension
gcc -pthread -B /pylon5/ci560jp/xhchen/anaconda3/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/pylon5/ci560jp/xhchen/anaconda3/lib/python3.7/site-packages/torch/include -I/pylon5/ci560jp/xhchen/anaconda3/lib/python3.7/site-packages/torch/include/torch/csrc/api/include -I/pylon5/ci560jp/xhchen/anaconda3/lib/python3.7/site-packages/torch/include/TH -I/pylon5/ci560jp/xhchen/anaconda3/lib/python3.7/site-packages/torch/include/THC -I/pylon5/ci560jp/xhchen/anaconda3/include/python3.7m -c sparse_coo_tensor.cpp -o build/temp.linux-x86_64-3.7/sparse_coo_tensor.o -lcusparse -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=sparse_coo_tensor_cpp -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++11
cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++
sparse_coo_tensor.cpp: In function ‘void spmm_gpu(const at::Tensor&, const at::Tensor&, const at::Tensor&, int32_t, int32_t, at::Tensor&, at::Tensor&)’:
sparse_coo_tensor.cpp:125:19: error: ‘THCState_getCurrentSparseHandle’ was not declared in this scope
auto handle = THCState_getCurrentSparseHandle(state);
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
sparse_coo_tensor.cpp:125:19: note: suggested alternative: ‘THCState_getCurrentStream’
auto handle = THCState_getCurrentSparseHandle(state);
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
THCState_getCurrentStream

Any idea what is wrong?

Best,
Xuhao Chen
https://chenxuhao.github.io/

Hi!
I'm trying to reproduce your brilliant works on our environment. However, we encounters several compilation problems due to incompatible dependency libraries. We are using CUDA11.4. It would be nice if you can give some suggestions on which branch of your codebase should be used for easier reproduction. So far we tried your newest branch trackml, but with compilation error as fatal error: nsparse_asm.h: No such file or directory