:You need to move the header and library files into your system's CUDA installation (usually located at /usr/local/cuda-11.2/ ). Run these commands with sudo :
: Ensure /usr/local/cuda/lib64 is in your LD_LIBRARY_PATH environment variable so your software can find the libraries. cudnn-11.2-linux-x64-v8.1.1.33.tgz
: This specific build is for CUDA 11.x. While cuDNN 8.x is generally compatible across CUDA 11.x versions, using the exact matching CUDA 11.2 toolkit is recommended for stability with frameworks like TensorFlow 2.6. :You need to move the header and library
You should see values representing , Minor 1 , and Patch 1 . Troubleshooting While cuDNN 8
cat /usr/local/cuda/include/cudnn_version.h | grep CUDNN_MAJOR -A 2 Use code with caution. Copied to clipboard
Downloading cuDNN Backend for Windows Tarball Installation (zip) - per-CUDA cuDNN versions are provided as separate tarballs (zip) NVIDIA Docs
This will create a directory named cuda containing include and lib64 subdirectories.