如何解决Pop!_OS系统中TensorFlow训练MNIST基础CNN模型时的cuDNN初始化错误？

Hey, I’ve helped several Pop!_OS users work through this exact issue—even when GPU detection looks normal, cuDNN initialization can fail for a handful of easy-to-miss reasons. Let’s walk through the most effective fixes:

1. Double-check version compatibility between TensorFlow, CUDA, and cuDNN #

This is the #1 culprit. TensorFlow has strict version locks for CUDA and cuDNN—even a minor version mismatch can break convolution operations.

• First, run this command to get your TensorFlow version:

  python -c "import tensorflow as tf; print(tf.__version__)"
  

• Then cross-reference that version against TensorFlow’s official compatibility table to make sure your installed CUDA and cuDNN versions are an exact match. For example, TensorFlow 2.15 requires CUDA 12.2 and cuDNN 8.9—no substitutions.

2. Verify cuDNN files are properly placed in CUDA directories #

Pop!_OS uses /usr/local/cuda as the default CUDA root directory. Make sure your cuDNN files are copied correctly here:

• After extracting the cuDNN zip archive, run these commands to move the files:

  sudo cp include/cudnn*.h /usr/local/cuda/include
  sudo cp lib64/libcudnn*.so* /usr/local/cuda/lib64
  

• Update the system library cache with:

  sudo ldconfig
  

This ensures the system can find the cuDNN libraries when TensorFlow calls them.

3. Fix missing or incorrect environment variables #

Sometimes the system can’t locate CUDA/cuDNN even if they’re installed because environment variables aren’t set. Add these lines to your ~/.bashrc (or ~/.zshrc if you use Zsh):

export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
export CUDA_HOME=/usr/local/cuda

Then reload the config file with source ~/.bashrc and restart your terminal before running your model again.

4. Free up GPU memory or enable memory growth #

Other processes (like desktop GPU acceleration, or leftover AI model sessions) can hog GPU memory and prevent cuDNN from initializing.

• Use nvidia-smi to check for running GPU processes. If you see any unexpected ones, kill them before launching your model.
• Alternatively, tell TensorFlow to only use GPU memory as needed (instead of grabbing all at once) by adding this code at the start of your script:

  import tensorflow as tf
  
  gpus = tf.config.list_physical_devices('GPU')
  if gpus:
      try:
          for gpu in gpus:
              tf.config.experimental.set_memory_growth(gpu, True)
          print("GPU memory growth enabled")
      except RuntimeError as e:
          print(e)
  

5. Do a clean reinstall of CUDA and cuDNN #

If all else fails, residual files from previous installs might be causing conflicts. First, completely remove existing CUDA installations:

sudo apt-get purge nvidia-cuda*
sudo rm -rf /usr/local/cuda*

Then follow TensorFlow’s installation guide step-by-step to install the exact compatible versions of CUDA and cuDNN for your TensorFlow release.

内容的提问来源于stack exchange，提问作者Pinkman8144