The following wheel has been compiled on xeon x5660 , a cpu with no AVX support.
** If you want to get the wheel by cloning the repo, make sure you have git lfs installed since this repo is using it !**, if not, please download it directly from github.
before installing it , follow these steps:
-
get the correct version for cuda to make it work :
-
sudo sh cuda_11.0.3_450.51.06_linux.run
-
Make sure the PATH is correctly edited after the install , if not , edit it correctly with regard to the output installation instructions.
PS: Just in case reboot the computer here , I am not sure it is mandatory, but who knows;)
-
get the correct version of cudnn :cudnn-11.3-linux-x64-v8.2.0.53 on nvida cudnn archive (Download cuDNN v8.2.0 (April 23rd, 2021), for CUDA 11.x then cuDNN Library for Linux (x86_64)) : https://developer.nvidia.com/rdp/cudnn-archive
-
copy the include folder's files of the downloaded cudnn files in /usr/local/cuda-11.0/include
-
do the same with the lib64 files. copy them in /usr/local/cuda-11.0/lib64
-
finally, execute : sudo chmod a+r /usr/local/cuda-11.0/include/cudnn*.h /usr/local/cuda-11.0/lib64/libcudnn*
again make a reboot here, just in case.
python3.8 -m pip install path/to/the/wheel/tensorflow-2.4.1-cp38-cp38-linux_x86_64.whl
-
import tensorflow as tf
-
physical_devices = tf.config.list_physical_devices()
-
print(physical_devices)
If everything works , you should have no core dump when importing tensorflow and the physical_devices variable should contain a GPU if your computer has a NVIDIA card.