Nvidia-docker: Publish Images Based on Ubuntu 18.04 LTS

[cancan101]

This is a ticket to start tracking building images based on Ubuntu 18.04 LTS.

1. Issue or feature description

It looks like 18.04 (official) docker images are being published: https://hub.docker.com/r/library/ubuntu/ so it should be possible to start building and testing against them.

[dwSun]

good

[flx42]

It's still too early on our side, but this should happen at the end of the month.

[pisymbol]

Please, I depend on this. NVidia's lack of support of 17.10 has really hurt (I've had to use 17.04 docker images up until now since the nvidia-XXX drivers have completely diverged).

[davidwessman]

What is the status on this now? Would be super useful :+1:

[dwSun]

any update？
when will cuda support 18.04 officially?

[flx42]

It will happen next week. Hopefully, early next week.

[flx42]

Bad news, it won't happen this week. The release of CUDA 9.2 has been delayed so we can't get 18.04 images yet. CUDA 9.1 is not compatible with 18.04.

[pisymbol]

That sucks. Got an ETA though? A month? Three months? etc.

[flx42]

~2 weeks, that's what I was told.

[mns6070]

@flx42 Are we still on track for delivery this week?

[dwSun]

CUDA 9.1 is compatible with ubuntu18.04. The ubuntu official repo already have cuda included.
Type those commands to install it:

# for cuda
sudo aptitude install nvidia-cuda-toolkit --without-recommends

# for cudnn
tar -zxf cudnn-9.1-linux-x64-v7.tgz
cd cuda
mkdir -p lib/x86_64-linux-gnu
mv lib64/* lib/x86_64-linux-gnu
rm -rvf lib64
sudo rsync -avp ./ /usr

For those who speak chinese, there is a tutorial:
http://www.cnblogs.com/dwsun/p/7767210.html
or this:
https://www.tinymind.cn/articles/163

Tested on my pc with a nv geforce 1060 card.

[flx42]

@dwSun this won't work with gcc 7.3.0 (shipped with bionic), you are limited to clang 3.X if I remember correctly.

[KiaraGrouwstra]

fwiw following @dwSun's instructions I get as far as installing all that then nvidia-docker (using readme instructions), but get a cuda error trying to run the actual readme example.

[dwSun]

Sorry for lack of details.
Tested with TensorFlow 1.5 (GPU, CUDA 9.1, Generic) from tinymind. And mxnet-cu91mkl-1.1.0 from pypi. Running directly on host.

CPP examples which need be built are NOT tested, on host or inside nvidia-docker.

nvidia-docker can be installed, without cuda on host. I am using nvidia/cuda:9.1-cudnn7-runtime. this image based on ubuntu:16.04.
nvidia-smi inside nvidia-docker works as expected. But mxnet-cu91mkl-1.1.0 inside nvidia-docker stuck at image preprocessing.

[flx42]

nvidia-docker can be installed, without cuda on host. I am using nvidia/cuda:9.1-cudnn7-runtime. this image based on ubuntu:16.04.

Yeah, that's the benefit of containers, you can still run ubuntu 16.04/14.04 or CentOS 6/7 containers on a Ubuntu 18.04 host, even if Ubuntu 18.04 doesn't have an official CUDA release right now.
This is not new :)

But mxnet-cu91mkl-1.1.0 inside nvidia-docker stuck at image preprocessing.

This is probably unrelated.

[yf225]

@flx42 Will CUDA 9.2 be supporting Ubuntu 18.04? It doesn't seem to be on the OS version list: https://developer.nvidia.com/cuda-downloads?target_os=Linux&target_arch=x86_64&target_distro=Ubuntu

[flx42]

It will be experimental support, using the ubuntu 17.10 repository.

[flx42]

Experimental images are out!

CUDA 9.2 (EXPERIMENTAL)

• 9.2-base-ubuntu18.04 (9.2/base/Dockerfile)
• 9.2-runtime-ubuntu18.04 (9.2/runtime/Dockerfile)
• 9.2-devel-ubuntu18.04 (9.2/devel/Dockerfile)
• 9.2-cudnn7-runtime-ubuntu18.04 (9.2/runtime/cudnn7/Dockerfile)
• 9.2-cudnn7-devel-ubuntu18.04 (9.2/devel/cudnn7/Dockerfile)

[cancan101]

Any plans to have an 18.04 deb repos? Right now it looks like the Dockerfile uses the 1710: https://gitlab.com/nvidia/cuda/blob/ubuntu18.04/9.2/base/Dockerfile#L7
repos list: http://developer.download.nvidia.com/compute/cuda/repos/

[flx42]

Ubuntu 18.04 will get its deb repo for the next CUDA release: 10.0
I don't know when that will be.

Let me know if there is any problem with the current image.

[sursu]

Why is it that every time Ubuntu gets updated I cannot just start using the latest version of CUDA on it? Or vice-versa.
Do the dependencies really change so much?

I imagine that the updates need to go through some tests. If there are changes in functionalities, naming they need to be examined and addressed.
Does testing take that much time?

Or, to put it differently: what does "EXPERIMENTAL" mean? What are the caveats?

[KiaraGrouwstra]

@sursu have you had issues with it? I went through the readme instructions and successfully ran its test.

[sursu]

I'll have to install Ubuntu first on my machine with NVIDIA GPU, which I thought I'll postpone untill 18.04 supports CUDA 9.2.

That was more of a general curiosity of mine.

[flx42]

Why is it that every time Ubuntu gets updated I cannot just start using the latest version of CUDA on it? Or vice-versa. Do the dependencies really change so much?

A new CUDA version will require a recent driver, it's as simple as that.

Or, to put it differently: what does "EXPERIMENTAL" mean? What are the caveats?

Testing wasn't done on 18.04, only on Ubuntu 17.10. That's why it's "experimental"

[mrjj]

@dwSun

For those who speak chinese,

Chinese bash is wonderfully readable :) Thank you, helped a lot!