TensorFlow GPU配置报错：找不到libnvidia-fatbinaryloader.so.396.26

Hey Juhua, I’ve dealt with this exact version mismatch issue before—let’s get your TensorFlow GPU setup back on track. The core problem here is that your system is searching for libnvidia-fatbinaryloader.so.396.26, but your installed NVIDIA 396 driver only includes the 396.24 version of this file. This usually pops up when your CUDA Toolkit expects a specific minor driver version that doesn’t align with what you’ve installed.

Here are the solutions you can try, ordered from quickest to most robust:

1. Create a Symbolic Link (Quick Fix) #

The fastest way to resolve this is to link the existing 396.24 file to the 396.26 name the system is looking for. Run these commands in your terminal:

# Create the symbolic link in the nvidia library directory
sudo ln -s /usr/lib/nvidia-396/libnvidia-fatbinaryloader.so.396.24 /usr/lib/nvidia-396/libnvidia-fatbinaryloader.so.396.26
# Update the system's library cache to recognize the new link
sudo ldconfig

To confirm the link was created successfully, run:

ls -l /usr/lib/nvidia-396/libnvidia-fatbinaryloader.so.396.26

You should see a line pointing to libnvidia-fatbinaryloader.so.396.24. Now try running your TensorFlow GPU task again—this should fix the missing file error.

2. Reinstall the Exact Driver Version (Cleaner Fix) #

If the symbolic link doesn’t work (or you prefer a fully compatible setup), you can install the 396.26 version of the NVIDIA driver to match what CUDA expects. Here’s how:

• First, remove your current 396.24 driver:

  sudo apt-get purge nvidia-396
  sudo apt-get autoremove
  

• Download the 396.26 driver package (make sure it’s compatible with your Tesla K80) from NVIDIA’s official archives.
• Temporarily disable your X server to avoid installation conflicts:

  sudo systemctl stop lightdm
  

• Make the downloaded .run file executable and install it:

  chmod +x NVIDIA-Linux-x86_64-396.26.run
  sudo ./NVIDIA-Linux-x86_64-396.26.run
  

• Follow the on-screen prompts, restart your system, and reconfigure CUDA/cuDNN if necessary.

Why This Happens #

CUDA Toolkit 9 has specific dependencies on minor versions of NVIDIA driver components. Even though you’re using the correct 396 major version, the mismatch between the minor versions (24 vs 26) causes the system to fail locating the required library. The symbolic link tricks the system into using the existing file, while reinstalling the exact driver version ensures full compatibility.

内容的提问来源于stack exchange，提问作者eigen