raise a error when running dgl_install.sh "CUB requires a definition of CUB_NS_QUALIFIER when CUB_NS_PREFIX/POSTFIX are defined"

[Gibyeng]

It seems that CUB_NS_QUALIFIER should be defined in the header file "dgl_cub.cuh", but after I add following into the header file,

#ifndef CUB_NS_QUALIFIER
#define CUB_NS_QUALIFIER ::cub
#endif

there are still errors about the namespace, I have no idea how to fix it.

[lixiaobai09]

Please give your evaluation environment. I guess your CUDA version is 11+, the CUB library is already integrated into CUDA 11. But our CUDA version is 10.1 for DGL v0.7.1 in our evaluation, the CUB needs to be imported separately. So the build environment is changed. If you want to reproduce our evaluation with DGL v0.7.1, I recommend the docker environment.

Or you can update the DGL version for CUDA 11+.

[Gibyeng]

@lixiaobai09 Thanks for your help. I install cuda 10.1 instead and it works! But when I try to import dgl, it fails again. There is the message I got.
’OSError: libcublas.so.11: cannot open shared object file: No such file or directory’.

[lixiaobai09]

This seems that you add the libcublas 11 in CUDA 11 to your link library path, but compile with CUDA 10. Maybe you can add the CUDA 10 library in the Linux environment variable LD_LIBRARY_PATH and rebuild DGL to solve it.

[Gibyeng]

@lixiaobai09 Thanks a lot! I installed dgl successfully.
I find "--num-train-worker" and "--num-sample-worker" are in the input arguments. It seems that FGNN requires two GPUs at least (one for training and one for sampling). Could I make it work with single GPU?
And when I run the code in the example i got this error:

[samgraph/common/common.cc:99]:Check failed: (data) != ((void *)-1)