Skip to content

Latest commit

 

History

History
 
 

gpu

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
aws
aws
 
 
gcp
gcp
 
 
docker-compose-2-gpus.yml
docker-compose-2-gpus.yml
 
 
docker-compose-4-gpus.yml
docker-compose-4-gpus.yml
 
 
docker-compose-gpu.yml
docker-compose-gpu.yml
 
 
fix_docker_compose.py
fix_docker_compose.py
 
 
generate_multi_gpu.py
generate_multi_gpu.py
 
 
install_cuda.sh
install_cuda.sh
 
 
install_docker.sh
install_docker.sh
 
 
install_nvidia_docker.sh
install_nvidia_docker.sh
 
 
readme.md
readme.md
 
 

readme.md

Deployment on single machines with multiple GPUs

The docker-compose files in this repo are intended for single instance GPU deployment. E.g. A single EC2 or a GCP instance. For several one-off workloads this is good enough.

  • install_cuda_drivers.sh install CUDA along with drivers on GCP Ubuntu Xenial VM ( Not need for AWS since the DL AMI contains pre-installed drivers.)

  • docker-compose-gpu.yml : Docker compose file for single GPU with at last 12 Gb VRAM.

  • docker-compose--gpus.yml : Docker compose file for multiple GPUs.

  • fix_docker_compose.py : make nvidia-docker default runtime.

  • install_docker.sh : install compatible docker version, make sure you log out and log in.

  • install_nvidia_docker.sh install nvidia docker and make it default runtime.

  • /aws instructions for using P2 instance and AMI

  • /gcp instructions for using Google Cloud Platform cloud VM

The docker-compose files use loop back interface (127.0.0.1:8000:80), we recommend forwarding the host OS port (8000)

over SSH tunnel when using cloud providers or VPS services such as Linode.