cudaErrorDriverLibraryNotLoaded
cuda
system_error
ai_generated
true
CUDA 错误:驱动程序库未加载 (cudaErrorDriverLibraryNotLoaded)
CUDA error: driver library is not loaded (cudaErrorDriverLibraryNotLoaded)
ID: cuda/driver-library-not-loaded
82%修复率
87%置信度
1证据数
2023-11-01首次发现
版本兼容性
| 版本 | 状态 | 引入 | 弃用 | 备注 |
|---|---|---|---|---|
| CUDA 11.8 | active | — | — | — |
| CUDA 12.2 | active | — | — | — |
| CUDA 12.4 | active | — | — | — |
根因分析
CUDA 驱动程序共享库 (libcuda.so) 未找到或无法加载,通常是因为未安装 NVIDIA 驱动程序或缺少库路径。
English
The CUDA driver shared library (libcuda.so) is not found or cannot be loaded, usually because the NVIDIA driver is not installed or the library path is missing.
官方文档
https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__DRIVER.html解决方案
-
Install the NVIDIA driver matching your GPU and CUDA version. For Ubuntu 22.04 with CUDA 12.2: sudo apt update; sudo apt install nvidia-driver-535; sudo reboot. Verify with 'nvidia-smi' and 'ldconfig -p | grep libcuda'.
-
If the driver is installed but libcuda.so is not in the linker path, add it explicitly: sudo ldconfig /usr/lib/x86_64-linux-gnu/nvidia/current; or set LD_LIBRARY_PATH to that directory: export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu/nvidia/current:$LD_LIBRARY_PATH.
无效尝试
常见但无效的做法:
-
80% 失败
Installing the CUDA toolkit without the NVIDIA driver is insufficient; the driver provides libcuda.so.
-
70% 失败
Setting LD_LIBRARY_PATH to a wrong directory (e.g., /usr/local/cuda/lib instead of lib64) does not help and may break other libraries.