This example shows how use dstack to spin up a Ray cluster and run Ray jobs on it.

First create a fleet for the Ray cluster. We'll use one instance for a master node and three instances for worker nodes:

type: fleet
name: ray-fleet
nodes: 4
placement: cluster
backends: [gcp]
resources: 
  cpu: 8..
  memory: 32GB..
  gpu: 1

dstack apply -f fleet.dstack.yaml

The following dstack task launches Ray master and worker nodes. dstack makes the Ray dashboard available at localhost:8265.

type: task
name: ray-cluster
nodes: 4
commands:
  - pip install -U "ray[default]"
  - >
    if [ $DSTACK_NODE_RANK = 0 ]; then 
      ray start --head --port=6379;
    else
      ray start --address=$DSTACK_MASTER_NODE_IP:6379
    fi
ports:
  - 8265 # ray dashboard port
resources:
  shm_size: 8GB

dstack apply -f cluster.dstack.yaml

Install Ray locally:

pip install ray

Now you can submit Ray jobs to the cluster available at localhost:8265:

RAY_ADDRESS='http://localhost:8265' ray job submit \
--working-dir . \
--runtime-env-json='{"pip": ["ray[train]", "torch", "torchvision", "tqdm", "filelock"]}' \
-- python tasks/pytorch-mnist.py

See more examples in the Ray docs.

Using Ray via dstack is a powerful way to get access to the rich Ray ecosystem while benefiting from dstack's provisioning capabilities.