【发布时间】：2020-12-04 05:55:01
【问题描述】：
我在 Ubuntu 18.04 上。并输出以下命令：
nvidia-smi
Fri Dec 4 11:35:09 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.87.01 Driver Version: 418.87.01 CUDA Version: 10.1 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Tesla M40 On | 00000000:00:08.0 Off | Off |
| N/A 59C P0 146W / 250W | 11724MiB / 12215MiB | 100% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
+-----------------------------------------------------------------------------+
nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2018 NVIDIA Corporation
Built on Sat_Aug_25_21:08:01_CDT_2018
Cuda compilation tools, release 10.0, V10.0.130
dpkg -l | grep cuda
ii cuda-command-line-tools-10-0 10.0.130-1 amd64 CUDA command-line tools
ii cuda-compat-10-0 410.104-1 amd64 CUDA Compatibility Platform
ii cuda-cublas-10-0 10.0.130-1 amd64 CUBLAS native runtime libraries
ii cuda-cudart-10-0 10.0.130-1 amd64 CUDA Runtime native Libraries
ii cuda-cudart-dev-10-0 10.0.130-1 amd64 CUDA Runtime native dev links, headers
ii cuda-cufft-10-0 10.0.130-1 amd64 CUFFT native runtime libraries
ii cuda-cuobjdump-10-0 10.0.130-1 amd64 CUDA cuobjdump
ii cuda-cupti-10-0 10.0.130-1 amd64 CUDA profiling tools interface.
ii cuda-curand-10-0 10.0.130-1 amd64 CURAND native runtime libraries
ii cuda-cusolver-10-0 10.0.130-1 amd64 CUDA solver native runtime libraries
ii cuda-cusparse-10-0 10.0.130-1 amd64 CUSPARSE native runtime libraries
ii cuda-driver-dev-10-0 10.0.130-1 amd64 CUDA Driver native dev stub library
ii cuda-gdb-10-0 10.0.130-1 amd64 CUDA-GDB
ii cuda-gpu-library-advisor-10-0 10.0.130-1 amd64 CUDA GPU Library Advisor.
ii cuda-license-10-0 10.0.130-1 amd64 CUDA licenses
ii cuda-memcheck-10-0 10.0.130-1 amd64 CUDA-MEMCHECK
ii cuda-misc-headers-10-0 10.0.130-1 amd64 CUDA miscellaneous headers
ii cuda-nvcc-10-0 10.0.130-1 amd64 CUDA nvcc
ii cuda-nvdisasm-10-0 10.0.130-1 amd64 CUDA disassembler
ii cuda-nvprof-10-0 10.0.130-1 amd64 CUDA Profiler tools
ii cuda-nvtx-10-0 10.0.130-1 amd64 NVIDIA Tools Extension
ii cuda-repo-ubuntu1804 10.1.243-1 amd64 cuda repository configuration files
ii libcudnn7 7.4.1.5-1+cuda10.0 amd64 cuDNN runtime libraries
ii libnvinfer5 5.0.2-1+cuda10.0 amd64 TensorRT runtime libraries
ii nvinfer-runtime-trt-repo-ubuntu1804-5.0.2-ga-cuda10.0 1-1 amd64 nvinfer-runtime-trt repository configuration files
所以我安装了 cuda10.0。我也设置了路径：
export CUDA_HOME=/usr/local/cuda
export PATH=${CUDA_HOME}/bin:${PATH}
export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:$LD_LIBRARY_PATH
但是为什么会出现这个错误呢？它寻找的是 cuda10.1，而不是 cuda10.0？
python3 -c 'import tensorflow as tf; print(tf.__version__)'
2020-12-04 11:37:43.929779: W tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could not load dynamic library 'libcudart.so.10.1'; dlerror: libcudart.so.10.1: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda/lib64:/usr/local/hadoop/lib/native:/usr/lib/jvm/java-8-openjdk-amd64/jre/lib/amd64/server:/usr/local/cuda/extras/CUPTI/lib64:/usr/local/nvidia/lib:/usr/local/nvidia/lib64
2020-12-04 11:37:43.929830: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.
2.3.1
【问题讨论】：
-
您必须使用构建 tensorflow 的 CUDA 版本。这是一个不可协商的要求。在这种情况下是 10.1
-
@talonmies 错误消息是否意味着 tensorflow 是针对 cuda 10.1 构建的？那么，如果我降级 tensorflow 版本，错误信息会消失吗？
-
正如我在评论中所说，是的，这意味着您的 tensorflow 是针对 10.1 构建的并且需要 10.1。至于降级，我不知道。如果您安装了 CUDA 10.0，则需要使用基于 CUDA 10.0 构建的版本。可能不是降级。那是你自己决定的。
-
我还建议将 CUDA 升级到 10.1 而不是降级 TF。除了当前的问题之外，它还会使您免于其他不兼容的情况。
-
或者，如果您无法更新 CUDA，您可以针对您的 CUDA 版本构建所需版本的 TF。但是@PoeDator 的建议仍然是最可取的。
标签： tensorflow nvidia