/usr/local/cuda-12.6/extras/demo_suite/deviceQuery Should show your GPU and Result = PASS . Download from NVIDIA cuDNN archive (requires login). Example for cuDNN 9.x compatible with CUDA 12.6:
Check current driver:
cd ~ cat > test.cu << EOF #include <stdio.h> int main() printf("CUDA version: %d\n", __CUDACC_VER__); return 0; nvidia cuda toolkit 12.6
tar -xvf cudnn-linux-x86_64-9.x.x.x_cuda12-archive.tar.xz sudo cp cudnn-*/include/cudnn*.h /usr/local/cuda-12.6/include/ sudo cp cudnn-*/lib/libcudnn* /usr/local/cuda-12.6/lib64/ sudo chmod a+r /usr/local/cuda-12.6/include/cudnn*.h /usr/local/cuda-12.6/lib64/libcudnn* | Issue | Solution | |-------|----------| | gcc version too high | Use export CC=gcc-12 CXX=g++-12 before nvcc | | Driver mismatch | Ensure driver ≥550.54.15 ( nvidia-smi top-right) | | nvcc not found | Re-check PATH ; logout/re-login | | Missing libcuda.so | Install driver properly or set LD_LIBRARY_PATH | | Kernel build fails | sudo apt install linux-headers-$(uname -r) | 9. Uninstall sudo /usr/local/cuda-12.6/bin/cuda-uninstaller sudo rm -rf /usr/local/cuda-12.6 Summary CUDA Toolkit 12.6 is stable and widely compatible. Use the runfile method to keep your existing driver intact. Always verify with nvcc --version and deviceQuery . For deep learning, pair with cuDNN 9.x and a framework built for CUDA 12.6. /usr/local/cuda-12