use nvidia-docker-compose launch a container, but exited soon - docker-compose

Question

My docker-compose.yml file :
version: '2'
services:
zl:
image: zl/caffe-torch-gpu:12.27
ports:
- "8801:8888"
- "6001:6008"
devices:
- /dev/nvidia0
volumes:
- ~/dl-data:/root/dl-data
After nvidia-docker-compose up -d the container launched, but exited soon.
But when I launch a container by nvidia-docker way, it worked well.
nvidia-docker run -itd -p 6008:6006 -p 8808:8888 -v `pwd`:/root/dl-data --name zl_test

Answer 1

You don't have to use nvidia-docker-compose.
By configuring the nvdia-docker plugin correctly you can just use docker-compose!
Via the nvidia docker git repo:
(can confirm it works for me)
Step 1:
Figure out nvidia driver version (it matters).
run:
nvidia-smi
output:
+---------------------------------------------------------------+
NVIDIA-SMI 367.57 Driver Version: 367.57
|-------------------------------+--------+----------------------+
Step 2:
create a docker volume that uses the nvidia-docker plugin must be done outside of compose as compose will mangle the volume name if it creates it.
docker volume create --name=nvidia_driver_367.57 -d nvidia-docker
Step 3
in the docker-compose.yml file:
version: '2'
volumes:
nvidia_driver_367.57: # same name as one created above
external: true #this will use the volume we created above
services:
cuda:
command: nvidia-smi
devices: #this is required
- /dev/nvidiactl
- /dev/nvidia-uvm
- /dev/nvidia0 #in general: /dev/nvidia# where # depends on which gpu card is wanted to be used
image: nvidia/cuda
volumes:
- nvidia_driver_367.57:/usr/local/nvidia/:ro