Install Issue: torch/_C.so looks for two versions of cuda

[wk910930]

Hi,

When I am trying to import torch, it said ImportError: libcudart.so.7.5: cannot open shared object file. However, when I run the command ldd /usr/local/lib/python2.7/dist-packages/torch/_C.so, it shows

...
linux-vdso.so.1 =>  (0x00007ffcc25f4000)
libshm.so => /usr/local/lib/python2.7/dist-packages/torch/lib/libshm.so (0x00007fa881131000)
libcudart.so.8.0 => /usr/local/cuda/lib64/libcudart.so.8.0 (0x00007fa880eca000)
libnccl.so.1 => /usr/local/lib/libnccl.so.1 (0x00007fa87eead000)
libTH.so.1 => /usr/local/lib/python2.7/dist-packages/torch/lib/libTH.so.1 (0x00007fa87e8e2000)
...
libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007fa86ed3f000)
librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007fa86eb37000)
libcudart.so.7.5 => not found
libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007fa86e830000)
libopenblas.so.0 => /usr/lib/libopenblas.so.0 (0x00007fa86d01d000)
...

Somehow it is looking for the libcudart of two different versions (7.5 & 8.0) at the same time! Currently in my machine, there is only cuda-8.0 installed and /usr/local/cuda is linking to cuda-8.0, too. I have been confused by this for 3 days.

Someone encountered the same issue?

Thanks!

[pavanky]

How did you install PyTorch?

[wk910930]

I tried sudo python setup.py install and CUDA_HOME=/usr/local/cuda-8.0 CUDA_BIN_PATH=/usr/local/cuda-8.0/bin CUDNN_LIB_DIR=/opt/library/NVIDIA/cudnn/v6.0_cuda_8.0/lib64 CUDNN_INCLUDE_DIR=/opt/library/NVIDIA/cudnn/v6.0_cuda_8.0/include sudo python setup.py install, none of them work.

[pavanky]

It is possible that some static library being linked with _C.so was linked with CUDA 7.5.

[wk910930]

Really? Is there any way I can clean/fix it?

[wk910930]

Problem Solved: Actually, it is the nccl (which was build with cuda-7.5) that is looking for libcudart.so.7.5. Just clean and rebuild nccl with cuda-8.0, and then sudo python setup.py install, now it works like a charm!

Thank you @pavanky !

Closing the issue.