CUDA Path Not Correctly Configured

1

Hello,
I am trying to configure Pytorch with CUDA support. However, torch.cuda.is_available() keeps on returning false.

Checking nvidia-smi, I am using CUDA 10.0. I modified my bash_profile to set a path to CUDA.

export PATH=/usr/local/cuda-10.0/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64\{LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

When attempting to use CUDA, I received this error.
THCudaCheck FAIL file=/pytorch/aten/src/THC/THCGeneral.cpp line=50 error=30 : unknown error

Is there some solution?

2

You can always try to set the environment variable CUDA_HOME. By the way, one easy way to check if torch is pointing to the right path is

from torch.utils.cpp_extension import CUDA_HOME
print(CUDA_HOME) # by default it is set to /usr/local/cuda/
3

Interestingly, I got no CUDA runtime found despite assigning it the CUDA path. nvcc did verify the CUDA version.

4

That’s weird.

CUDA_HOME=a/b/c python -c "from torch.utils.cpp_extension import CUDA_HOME; print(CUDA_HOME)"

This prints a/b/c for me, showing that torch has correctly set the CUDA_HOME env variable to the value assigned.

5

It detected the path, but it said it can’t find a cuda runtime.

No CUDA runtime is found, using CUDA_HOME='/usr/local/cuda-10.0'

6

excuse me did you solve it ?

7

Did you find the solution to this issue? I have the same problem. Thanks