How to use GPU with stack, docker, nvidia-docker #157
Conversation
|
Can one of the admins verify this patch? |
|
Jenkins, test this please. |
README.md
Outdated
| @@ -82,6 +82,17 @@ There is also a demo application: | |||
| cd tensorflow-mnist | |||
| stack --docker --docker-image=$IMAGE_NAME build --exec Main | |||
|
|
|||
| If you want to use GPU you can do: | |||
There was a problem hiding this comment.
Maybe add a header like
### Docker GPU Supportso that it easier to scan through the docs.
README.md
Outdated
|
|
||
| We need stack to use nvidia-docker by using a 'docker' wrapper script. | ||
|
|
||
| ln -s `pwd`/tools/nvidia-docker-wrapper.sh <somewhere in your path>/docker |
There was a problem hiding this comment.
I think it is worth including a note that this will shadow the normal docker command.
Filed commercialhaskell/stack#3499 to ask for a nicer way to do this.
(Also, interestingly, I stumbled upon a blog in Japanese where someone documented another workaround, but this one seems simpler: https://shigekikarita.github.io/blog/2017/07/08/023.html)
There was a problem hiding this comment.
I'll add a note. I have heard rumours that nvidia-docker v2 will be integrated into docker proper. I don't have a source for this. At least a way to invoke it nicely... so we won't even need the nvidia-docker wrapper. Crossing fingers :)
docker/gpu/Dockerfile
Outdated
| # Support output of Unicode characters. | ||
| RUN dpkg-reconfigure locales && \ | ||
| locale-gen en_US.UTF-8 && \ | ||
| update-locale LANG=en_US.UTF-8 |
There was a problem hiding this comment.
Can you fix the indentation here and above?
There was a problem hiding this comment.
Will do!
I am in the process of testing your upgrade (#161) to 1.3 now, and will update this PR as necessary.
jcberentsen
force-pushed
the
dockergpu
branch
from
4ad30ba to
9765220
Compare
|
Jenkins, test this please. |
|
The changes LGTM, let me know when you think it is ready to merge. Thanks! |
jcberentsen
force-pushed
the
dockergpu
branch
from
9765220 to
5c28a58
Compare
|
I think it is ready to merge now. I got it working with tensorflow-gpu:1.3.0. |
|
Jenkins, test this please. |
| @@ -1,6 +1,6 @@ | |||
| # Prepare the image with: | |||
| # docker build -t tensorflow/haskell:v0 docker | |||
| FROM tensorflow/tensorflow:1.0.0 | |||
| FROM tensorflow/tensorflow:1.3.0 | |||
There was a problem hiding this comment.
Ah, I missed this when doing the update. Did using the old image cause any problems for you? It seems to work fine for me with 1.0.0 (but we should update it of course).
This procedure is a bit convoluted, but possible to get utilisation of GPU this way.