简体繁体 English

Tensorflow 2.3 未找到 GPU

[英]Tensorflow 2.3 doesn't find GPU

原文 2020-12-10 01:21:10 1 1 tensorflow/ installation/ gpu

For some reason my GPU is not used by tensorflow.由于某种原因，tensorflow 不使用我的 GPU。 When I run tf.test.is_built_with_cuda() it returns False .当我运行tf.test.is_built_with_cuda()它返回False 。 I have tensorflow-gpu installed, version 2.3.我安装了 tensorflow-gpu，版本 2.3。 Also I installed CUDA 10.1 along with cuDNN 7.65 and have the correct environment variables.我还安装了 CUDA 10.1 和 cuDNN 7.65 并具有正确的环境变量。

When I ran nvcc -V in the console it verifies that my CUDA is release 10.1, V10.1.243当我在控制台中运行nvcc -V时，它会验证我的 CUDA 是版本 10.1，V10.1.243

When I ran nvidia-smi it shows that my Driver version is 460.79 and the CUDA version is 11.2当我运行nvidia-smi时，它显示我的驱动程序版本是 460.79，CUDA 版本是 11.2

Do I maybe need to downgrade my GPU driver?我是否可能需要降级我的 GPU 驱动程序？

1 个解决方案

I too have the same problem, but when I installed PyTorch it correctly detected my GPU, previously tf 2.1.0 was working fine.我也有同样的问题，但是当我安装 PyTorch 时，它正确地检测到了我的 GPU，之前的tf 2.1.0工作正常。

Maybe it's a bug.也许这是一个错误。

I will try to install tf 2.1 and see whether it detects the GPU or not.我将尝试安装tf 2.1并查看它是否检测到 GPU。

UPDATE 1/7/2021: Tried installing the tf 2.1.0 version through conda installation but it didn't work.更新 1/7/2021：尝试通过 conda 安装安装 tf 2.1.0 版本，但没有成功。

Tensorflow was able to detect my gpu.I uninstalled the tf from conda installation and installed it through pip .I think the conda version doesn't detect the gpu.Make sure you the tensorflow version you are installing matches the requirement of the CUDA and cudnn library. Tensorflow was able to detect my gpu.I uninstalled the tf from conda installation and installed it through pip .I think the conda version doesn't detect the gpu.Make sure you the tensorflow version you are installing matches the requirement of the CUDA and cudnn图书馆。

My current configuration: tensorflow-gpu : 2.3.0我当前的配置： tensorflow-gpu ：2.3.0

CUDA :7.6.5 CUDA ：7.6.5

cudnn :10.1 cudnn :10.1