Initial attempts at installing via either the command line installer or rpm would fail with graphics driver error not installed due to lack of kernel-sources. (they are present and didn’t work when forced with --kernel-source-path=)
./cuda_9.0.103_384.59_linux.run --kernel-source-path /usr/src/kernels/4.11.12-200.fc25.x86_64/kernel/
Would fail
- the previous driver would build fine.
The only way I could make progress was to extract the packages:
./cuda_9.0.103_384.59_linux.run --extract /root/cudaextract
Then manually install the driver:
./NVIDIA-Linux-x86_64-384.59.run
followed by a telinit 5 to bring up the graphical desktop to check its working and telinit 3 to go back to the console.
Next I installed Cuda 9 - but not from the actual extracted package this gives a checksum error (yet the whole cuda_9.0.103_… has a valid md5sum???) - maybe that’s the problem?
Cuda 9 is installed via,
./cuda-linux.9.0.103-22377686.run
use n for the driver install (its already installed) and as you wish for other options.
Cuda 9 installs and promptly says its not. - I think this is in error since at least,
"/usr/local/cuda-9.0"/bin/nvcc -ccbin clang -I/home/lab/cuda9samples/NVIDIA_CUDA-9.0_Samples/common/inc/ -m64 --ptxas-options=-v -gencode arch=compute_61,code=sm_61 -gencode arch=compute_61,code=compute_61 -o test-cuda.o -c test-cuda.cu ; "/usr/local/cuda-9.0"/bin/nvcc -ccbin clang -m64 --ptxas-options=-v -gencode arch=compute_61,code=sm_61 -gencode arch=compute_61,code=compute_61 -std=c++11 -o test-cuda test-cuda.o -lcufft -lpthread
Resulted in a binary that worked.
Hope this is of some help to others trying the release candidate.
-Mike.