CUDA 10.1 + Tensorflow 1.13

The recently released TF 1.13 is built against CUDA 10 (.0 presumably?)

For a new build with TF 1.13 as the target application, is it recomennded to use CUDA 10.0 or 10.1?

10.0 for sure. tensorflow seems to be a fragile piece of software, everytime there is a cuda update it breaks. At this point apparently only the latest TF 1.13.1 released less than a week ago compiles with cuda 10.0, doubt that any tensorflow in release would work with 10.1, besides cuda 10.1 seems to be broken for other reason, see other threads.

@bernard Yes Tensorflow 1.13.1 dosen’t work with CUDA 10.1. I am in process to compile 1.13.1 with CUDA 10.0. Hope it works

@bernard Tensorflow 1.13.1 worked with CUDA 10.0, cudnn 7.3. Tested successfully

I dont recommend tf 1.13 https://github.com/tensorflow/tensorflow/issues/26218

Yes Tensorflow 1.13.1 dosen’t work with CUDA 10.1.

How to down-grade CUDA 10.1 to CUDA 10.0?

Use the script generated by CUDA installer as

sudo /usr/local/cuda-X.Y/bin/uninstall_cuda_X.Y.p

Replace XY with your version of CUDA.

@a.khan

How to down-grade CUDA 10.1 to CUDA 10.0?

What is your OS and how did you install cuda?

If you are on Ubuntu and installed cuda with Nvidia’s ubuntu repo then it is really simple

sudo apt uninstall cuda
sudo apt autoremove
sudo apt update
sudo apt install cuda-10-0

Here “cuda” and “cuda-10-0” are metapackages cuda always update to the latest, in this case cuda10.1. Whereas cuda-version-number doesn’t update (others available are cuda-9-2, cuda-9-1 etc, you can easily search for them if you have synaptic, or use the command “apt-cache search cuda”)

If you are on Windows or Mac then I have no idea.

From the documentation

Ubuntu
To remove CUDA Toolkit:
$ sudo apt-get --purge remove "*cublas*" "cuda*"
To remove NVIDIA Drivers:
$ sudo apt-get --purge remove "*nvidia*"

Though it defines different uninstall for runfile and debfile installations at some other paragraph.
However, once removed - reboot and install cuda 10.0
it can be downloaded from https://developer.nvidia.com/cuda-10.0-download-archive

Though it defines different uninstall for runfile and debfile installations at some other paragraph.
However, once removed - reboot and install cuda 10.0

You don’t need to download from archive again. Nvidia’s deb repository has multiple cuda

All you have to do, instead of installing cuda

sudo apt install cuda-10-0

They seem to have taken the cuda-10-0 target out of the latest repo.

cuda-repo-ubuntu1810/unknown,now 10.1.105-1 amd64 [installed]
cuda repository configuration files

greg@salt:~$ sudo apt list cuda
Listing… Done
cuda/unknown,now 10.1.105-1 amd64 [installed]

The network .deb file from the 10.0 archive (https://developer.nvidia.com/cuda-10.0-download-archive?target_os=Linux&target_arch=x86_64&target_distro=Ubuntu&target_version=1804&target_type=debnetwork) should get you what you need. I just tried it, and I was able to run sudo apt-get install cuda-10-0 after uninstalling cuda 10.1.

TF 1.1.4 has been released!

Anybody know which Cuda it has been built against? No mention in the release notes.

I would also like to know.

TensorFlow 1.14 only works with CUDA 10.0, as I recently learned the hard way.

This shows the dependency on CUDA version 10.0 as well: https://www.tensorflow.org/install/gpu