Build for cuda 10 / RTX 20xx

[lahwran]

Hey! I'm looking into what it would take to build deepo for cuda 10. I'm creating this issue as a somewhat of a placeholder, I'll put more details in the comments as I work through it. Is anyone already working on this?

Context: I'm one of the founders of vast.ai, a service that allows renting others' gpus, and we use docker; Deepo is one of our main suggested images. We have some folks with 2080 machines, but we're blocked on having image support for cuda 10 builds of the frameworks - eg, cuda 9 tf seems to break on cuda 10.

Thanks for the awesome images, btw :)

[lahwran]

we've got a tensorflow image working separate from deepo, so I'm pausing on this for now.

[sanzgiri]

I am also interested in a build with Cuda 10 and latest CuDNN. Does anyone have a dockerfile for this?

[emican86]

sanzgiri, here is what I was able to find:
nvidia/cuda has an 16.04 CUDA 10 devel image availble

• 10.0-cudnn7-devel-ubuntu16.04 10.0/devel/cudnn7/Dockerfile

From tensorflow:
tensorflow/tensorflow#22706

• Nightly Builds are now CUDA 10 as of 16-DEC-2018
  Partial dockerfiles and apt commands:
• Devel Confirmed by building TensorFlow at 1.12RC0.
• Runtime
• Install Tensorflow w/ GPU & CUDA 10 from scratch

[sanzgiri]

@emican86 - I was able to do a cuda 10 build by changing the first line in Dockerfile.all-jupyter-py36-cu90 to
FROM nvidia/cuda:10.0-cudnn7-devel-ubuntu16.04

Also had to update the nvidia drivers to 410+ which I did using the instructions here: https://www.pytorials.com/how-to-install-tensorflow-gpu-with-cuda-10-0-for-python-on-ubuntu/ (this has been my go-to guide for cuda, cudnn installs).

[ufoym]

We do have a ubuntu18-py37-cu100 branch.
However, it fails to build torch because of several non-trivial issues.
So we are not going to provide CUDA10-based images
unless there are no longer any problem with all popular frameworks.

Alternatively, one can try --cuda-ver 10.0 for a specific framework, e.g,

python generator/generate.py Dockerfile tensorflow python==3.6 --cuda-ver 10.0 --cudnn-ver 7