Skip to content

Latest commit

 

History

History
111 lines (85 loc) · 3.91 KB

README.md

File metadata and controls

111 lines (85 loc) · 3.91 KB

docker-colab-local

Docker Cloud Build Status

A Docker container to act as a local runtime for Google Colab or private jupyter server with BERT preinstalled.

Get the Image

docker pull sorokine/docker-colab-local:latest

docker run can download the image automatically but docker pull will ensure that you have the latest build.

Run

docker run \
  --runtime=nvidia \
  -it --rm -p 8081:8081 \
  sorokine/docker-colab-local:latest

Or, to mount a volume so that it's accessible to colab:

docker run \
  --runtime=nvidia \
  -it --rm -p 8081:8081 \
  -v "$PWD":/opt/colab \
  sorokine/docker-colab-local:latest

where "$PWD" is a full path on your host machine to put your notebooks (other than the ones hosted on colab) and other files created by the notebooks (e.g., computed indices).

To make downloaded BERT models persistent run as:

docker run \
  --runtime=nvidia \
  -it --rm -p 8081:8081 \
  -v "$PWD":/opt/colab \
  -v $HOME/.cache/torch:/root/.cache/torch \
  sorokine/docker-colab-local:latest

This will use the same cache directory as the models run on the host (if your host is Linux or Mac).

Making Nvidia Work

  • on some docker versions replace --runtime=nvidia with --gpus all.
  • if you do not have a GPU or Nvidia drivers installed completely omit all nvidia flags from docker command. The notebook will work but it will be slow.
  • if you do not have the latest CUDA installed on your system check which tags are available in docker hub repo and use the one for your version of CUDA, e.g.:
docker run \
  --runtime=nvidia \
  -it --rm -p 8081:8081 \
  -v "$PWD":/opt/colab \
  -v $HOME/.cache/torch:/root/.cache/torch \
  sorokine/docker-colab-local:10.1
  • To find the version of CUDA on your system run nvidia-smi.

Connecting

If the container isn't running on your local machine, you'll need to forward port 8081. Run this command from the system where you are runing your browser:

ssh MACHINE_WHERE_DOCKER_IS_RUNNING -L 8081:localhost:8081

In Colaboratory, click the "Connect" button and select "Connect to local runtime...". Enter the port 8081 step in the dialog that appears and click the "Connect" button. (from colaboratory). Only localhost hostname is accepted (no numeric IPs). Replace the token in the dialog box with the token that is shown in the terminal after starting docker container. The connection string should look like http://localhost:8081/?token=abcdef123456.....

If your recieve an error like 0.0.0.0:8081 failed: port is already allocated.. try running container on a diferent port, for example for port 8082:

docker run \
  --runtime=nvidia \
  -it --rm -p 8082:8081 \
  -v "$PWD":/opt/colab \
  -v $HOME/.cache/torch:/root/.cache/torch \
  sorokine/docker-colab-local:latest

Colab notebooks expect the port in connection and on the server to be the same. To fix this problem without rebuilding the image forward the port with ssh:

$ ssh MACHINE_WHERE_DOCKER_IS_RUNNING -L 8081:localhost:8082

Running on a Remote System

This is a single command to run the image on a remote system with the GPU enabled and using the specified port:

( COLAB_PORT=8082; \
  ssh REMOTE_HOST_NAME -t -L $COLAB_PORT:localhost:$COLAB_PORT \
  docker run \
  --gpus=all \
  -p $COLAB_PORT:$COLAB_PORT \
  -v "$PWD":/opt/colab \
  -v $HOME/.cache/torch:/root/.cache/torch \
  --env COLAB_PORT=$COLAB_PORT \
  -it --rm \
  sorokine/docker-colab-local:10.1 \
)

Notes

  • if some packages are missing install them with !pip install in your notebook. This has to be repeated on kernel restart.

TODO

plans and pressing ussues will go here