TensorFlow无法dlopen GPU库且无具体警告的问题求助

问题现象 #

新装Linux Mint 21.1后，运行TensorFlow无法识别GPU，执行命令：

$ python3 -c "import tensorflow as tf;tf.config.list_physical_devices()"

得到输出：

2023-04-16 00:25:24.827994: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable TF_ENABLE_ONEDNN_OPTS=0.
2023-04-16 00:25:25.073486: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-04-16 00:25:26.881014: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:996] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
2023-04-16 00:25:26.881251: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1956] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu for how to download and setup the required libraries for your platform.
Skipping registering GPU devices...

注：通常这类错误会提示具体缺失库，但本次无相关提示；nvidia-smi显示正常，PyTorch可正常使用GPU

环境信息 #

• GPU：ASUS ROG STRIX RTX 3090
• Python版本：3.10.6
• TensorFlow版本：2.12.0
• CUDA版本：11.8.0-1
• CuDNN版本：8.6.0.163-1+cuda11.8
• TensorRT版本：8.6.0.12-1+cuda11.8
• NVIDIA驱动版本：515.105.01（曾尝试520、525版本，问题依旧）

解决方法 #

1. 修正CuDNN版本兼容性 #

TensorFlow 2.12.0官方要求搭配CuDNN 8.9.x（对应CUDA 11.8），当前使用的8.6.0版本不匹配，这是核心问题。

2. 配置环境变量 #

确保CUDA和CuDNN的库路径被系统正确识别：
编辑~/.bashrc（或~/.zshrc，根据你使用的shell），添加以下内容：

export LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/include:$LD_LIBRARY_PATH
export PATH=/usr/local/cuda/bin:$PATH

执行source ~/.bashrc使配置立即生效。

3. 验证库加载状态 #

用ldd检查TensorFlow的CUDA依赖是否能正常加载：

ldd $(python3 -c "import tensorflow as tf; print(tf.__file__)") | grep cuda

若输出中有缺失的库，根据提示安装对应包或调整路径。

4. 重装适配版本的CuDNN #

卸载现有CuDNN，安装TensorFlow要求的版本：

# 卸载旧版本
sudo apt remove libcudnn8 libcudnn8-dev
# 安装适配版本（以8.9.2为例）
sudo dpkg -i libcudnn8_8.9.2.26-1+cuda11.8_amd64.deb
sudo dpkg -i libcudnn8-dev_8.9.2.26-1+cuda11.8_amd64.deb

5. 测试GPU识别 #

重新运行测试命令：

python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

若输出包含GPU设备信息，说明问题已解决。

内容的提问来源于stack exchange，提问作者Regic