This guide will walk you through:
Defining a task in a simple YAML format
Provisioning a cluster and running a task
Using the core SkyPilot CLI commands
Be sure to complete the installation instructions first before continuing with this guide.
Let’s define our very first task, a simple Hello, SkyPilot! program.
Create a directory from anywhere on your machine:
$ mkdir hello-sky $ cd hello-sky
Copy the following YAML into a
resources: # Optional; if left out, automatically pick the cheapest cloud. cloud: aws # 1x NVIDIA V100 GPU accelerators: V100:1 # Working directory (optional) containing the project codebase. # Its contents are synced to ~/sky_workdir/ on the cluster. workdir: . # Typical use: pip install -r requirements.txt # Invoked under the workdir (i.e., can use its files). setup: | echo "Running setup." # Typical use: make use of resources, such as running training. # Invoked under the workdir (i.e., can use its files). run: | echo "Hello, SkyPilot!" conda env list
This defines a task with the following components:
resources: cloud resources the task must be run on (e.g., accelerators, instance type, etc.)
workdir: the working directory containing project code that will be synced to the provisioned instance(s)
setup: commands that must be run before the task is executed (invoked under workdir)
run: commands that run the actual task (invoked under workdir)
All these fields are optional.
To launch a cluster and run a task, use
$ sky launch -c mycluster hello_sky.yaml
This may take a few minutes for the first run. Feel free to read ahead on this guide.
You can use the
-c flag to give the cluster an easy-to-remember name. If not specified, a name is autogenerated.
sky launch command performs much heavy-lifting:
selects an appropriate cloud and VM based on the specified resource constraints;
provisions (or reuses) a cluster on that cloud;
syncs up the
In a few minutes, the cluster will finish provisioning and the task will be executed.
The outputs will show
Hello, SkyPilot! and the list of installed Conda environments.
Execute a task on an existing cluster¶
Once you have an existing cluster, use
sky exec to execute a task on it:
$ sky exec mycluster hello_sky.yaml
sky exec command is more lightweight; it
syncs up the
workdir(so that the task may use updated code); and
setup commands are skipped.
Bash commands are also supported, such as:
$ sky exec mycluster python train_cpu.py $ sky exec mycluster --gpus=V100:1 python train_gpu.py
For interactive/monitoring commands, such as
gpustat -i, use
ssh instead (see below) to avoid job submission overheads.
View all clusters¶
sky status to see all clusters (across regions and clouds) in a single table:
$ sky status
This may show multiple clusters, if you have created several:
NAME LAUNCHED RESOURCES COMMAND STATUS mygcp 1 day ago 1x GCP(n1-highmem-8) sky launch -c mygcp --cloud gcp STOPPED mycluster 4 mins ago 1x AWS(p3.2xlarge) sky exec mycluster hello_sky.yaml UP
See here for a list of all possible cluster states.
SSH into clusters¶
ssh <cluster_name> to log into a cluster:
$ ssh mycluster
Multi-node clusters work too:
# Assuming 3 nodes. # Head node. $ ssh mycluster # Worker nodes. $ ssh mycluster-worker1 $ ssh mycluster-worker2
The above are achieved by adding appropriate entries to
Because SkyPilot exposes SSH access to clusters, this means clusters can be easily used inside tools such as Visual Studio Code Remote.
After a task’s execution, use
scp to download files (e.g., checkpoints):
$ rsync -Pavz mycluster:/remote/source /local/dest # copy from remote VM
For uploading files to the cluster, see Syncing Code and Artifacts.
Stop/terminate a cluster¶
When you are done, stop the cluster with
$ sky stop mycluster
To terminate a cluster instead, run
$ sky down mycluster
Stopping a cluster does not lose data on the attached disks (billing for the instances will stop while the disks will still be charged). Those disks will be reattached when restarting the cluster.
Terminating a cluster will delete all associated resources (all billing stops), and any data on the attached disks will be lost. Terminated clusters cannot be restarted.
Find more commands that manage the lifecycle of clusters in the CLI reference.
So far, we have used SkyPilot’s CLI to submit work to and interact with a single cluster. When you are ready to scale out (e.g., run 10s or 100s of jobs), SkyPilot supports two options:
Queue many jobs on your cluster(s) with
sky exec(see Job Queue);
Use Managed Spot Jobs to run on auto-managed spot instances (users need not interact with the underlying clusters)
Managed spot jobs run on much cheaper spot instances, with automatic preemption recovery. Try it out with:
$ sky spot launch hello_sky.yaml
Congratulations! In this quickstart, you have launched a cluster, run a task, and interacted with SkyPilot’s CLI.
Adapt Tutorial: DNN Training to start running your own project on SkyPilot!
To learn more, try out SkyPilot Tutorials in Jupyter notebooks
We invite you to explore SkyPilot’s unique features in the rest of the documentation.