Win10 CUDA CUDNN installation configuration (torch paddlepaddle)

Foreword

Finally, the configuration of CUDA CUDNN is done this time. My graphics card is Geforece Nvidia 930MX.
This time record the configuration process.The cuda10.2 version has good support for torch and paddle, so this time we install cuda10.2

One: Check the driver that supports cuda10.2

https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html

cuda 10.2 driver version greater than or equal to 441.22

Two: Get a driver that supports your own GPU

https://www.nvidia.cn/geforce/drivers/

According to the first step, the driver we need to download is greater than or equal to 441.22

Enter the model number

Search for the driver that meets the requirements

Download and install

download cuda10.2 and cudnn

You can get it on the official website below, or you can get it directly from Baidu Netdisk
Link: https://pan.baidu.com/s/1W5IjQWDrT0kpmI1fMoapmQ?pwd=if33
Extraction code: if33

cuda 10.2 download address

https://developer.nvidia.com/cuda-10.2-download-archive?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exelocal

cudnn version download address

https://developer.nvidia.com/rdp/cudnn-archive

Unzip and install cuda10.2

Update.Core installation may report errors, just delete it

Select custom installation during installation, and then choose the installation path by yourself, of course, it is best to default directly

Unzip and add cudnn

Move the files in the folder corresponding to the unzipped cudnn to the folder corresponding to the cuda we installed

Add environment variables

Third test installation

nvcc -V

torch history

https://pytorch.org/get-started/previous-versions/

torch paddle tensorflow test whether the installation is successful

import tensorflow as tfprint(tf.test.is_gpu_available())print(tf.config.list_physical_devices('GPU'))import torchprint(torch.cuda.is_available())print(torch.Tensor(5, 3).cuda())import paddlepaddle.fluid.install_check.run_check()paddle.fluid.is_compiled_with_cuda()

Concluding remarks

This is almost done.