quickstart

Running MLX with Docker Compose

This "quickstart" setup uses Docker Compose to bring up the MLX API server and the MLX Dashboard, together with the MySQL database and Minio S3 storage backend. In this configuration, the preloaded asset catalog of Components, Datasets, Models, Notebooks and Pipelines can be browsed, asset metadata and sample code can be downloaded and new assets can be registered. Sample pipeline code can be generated for each asset type, however their execution on a Kubeflow Pipelines (KFP) cluster is not enabled. In a Kubernetes cluster deployment of MLX, Pipelines are registered using the KFP API and metadata storage is managed by KFP. In this Docker Compose setup the Pipelines are stored in Minio and MySQL by the MLX API server.

Limitations

The Kubeflow Pipelines dashboard and Inference Service capabilities are not available with this Docker Compose setup.

Prerequisites

• Install Docker Compose.
• It may be necessary to increase the Docker resources from the default of 2 GB memory to 4 GB.
• Approximately 10 GB of free storage

Clone this repository and navigate to the quickstart folder:

git clone https://github.com/machine-learning-exchange/mlx.git
cd mlx/quickstart

Keep up to date

If some time has passed since the mlx repository was cloned, make sure to pull the latest sources for the Quickstart:

git pull

Pull the Docker Images

Our Docker images for the mlx-api and mlx-ui get rebuilt nightly. To get the latest version, run:

docker compose pull

Bring up the Docker Containers

docker compose up

Wait for the containers to start up. When the MLX API and UI are ready, this message should show up in the terminal log:

dashboard_1   |
dashboard_1   | ================================================
dashboard_1   |  Open the MLX Dashboard at http://localhost:80/
dashboard_1   | ================================================
dashboard_1   |

Now open a web browser and type localhost in the address bar to open the MLX dashboard.

The MLX API spec can be explored at localhost:8080/apis/v1alpha1/ui/

Note: If the Docker compose stack is running on a remote host, and the MLX Web UI is running on localhost, export the environment variable DOCKER_HOST_IP, so that the MLX UI web app on localhost can connect to the MLX API on the Docker host.

export DOCKER_HOST_IP=127.0.0.1
docker compose up

Shut Down the Docker Containers

Press control + c on the Terminal to stop and then remove the containers:

docker compose down -v

Remove the Data Created by Minio and MySQL

docker volume prune -f

Troubleshooting

If you are working on a local clone of your fork, rather than a clone of the source repository, make sure to keep your code up to date:

git remote add upstream https://github.com/machine-learning-exchange/mlx.git
git fetch upstream
git checkout main
git rebase upstream/main
git push origin main --force

Since we are actively developing MLX, there may have been changes to the data schema which could conflict with the data created by running the Quickstart in days prior. The symptoms of this could be empty dashboards with endlessly spinning wheels. To remove all previously created Docker Compose data run the following commands:

docker compose down -v --remove-orphans
docker compose rm -v -f
docker volume prune -f

Windows Subsystem for Linux (WSL) Issues

Featured Assets Pages are Empty

If there are no featured asset cards showing up in the MLX web UI and the Docker Compose log shows an error like this:

catalog_1     | /bin/sh: /init_catalog.sh: not found
catalog_1 exited with code 127

Make sure you originally cloned/forked the source repo from inside the WSL sub-system, not Windows. This error happens because the MLX source files have Windows line endings (\r\n - CRLF) which bash cannot run. (https://askubuntu.com/questions/966488/how-do-i-fix-r-command-not-found-errors-running-bash-scripts-in-wsl#comment1553686_966488). This error in the catalog_1 service prevents the loading of assets and objects into the MLX catalog.