This quickstart deploys a barebone GPU-accelerated workload, a Python container that prints the result of nvidia-smi. It illustrates the basic components of a single run and how you can deploy one.

​
What you will do

  • Launch a GPU-accelerated workload
  • Set up a runtime for your model
  • Run a simple command: nvidia-smi

​
Installation & setup

After creating a free account at VESSL AI, install our Python package and get an API authentication. Set your primary Organization and Project for your account and let’s get going.

pip install --upgrade vessl
vessl configure
If you encounter the error ModuleNotFoundError: No module named 'packaging', please run the command pip install packaging.

​
Writing the YAML

Launching a workload on VESSL AI begins with writing a YAML file. Our quickstart YAML is in four parts:

  • Compute resource — typically in terms of GPUs — this is defined under resources
  • Runtime environment that points to a Docker Image — this is defined under image
  • Run commands executed inside the workload as defined under run

Let’s start by creating quickstart.yaml and define the key-value pairs one by one.

1

Spin up a compute instance

resources defines the hardware specs you will use for your run. Here’s an example that uses our managed cloud to launch an A10 instance.

You can see the full list of compute options and their string values for preset under Resources. Later, you will be able to add and launch workloads on your private cloud or on-premise clusters simply by changing the value for cluster.

resources:
  cluster: vessl-oci-sanjose
  preset: gpu-a10-small
2

Configure a runtime environment

VESSL AI uses Docker images to define a runtime. We provide a set of base images as a starting point but you can also bring your own custom Docker images by referencing hosted images on Docker Hub or Red Hat Quay.

You can find the full list of images and the dependencies for the latest build here. Here, we’ll use the latest go-to PyTorch container from NVIDIA NGC.

resources:
  cluster: vessl-oci-sanjose
  preset: gpu-a10-small
image: quay.io/vessl-ai/torch:2.3.1-cuda12.1
3

Write a run command

Now that we defined the specifications of the compute resource and the runtime environment for our workload, let’s run nvidia-smi.

We can do this by defining a pair of workdir and command under run.

resources:
  cluster: vessl-oci-sanjose
  preset: gpu-a10-small
image: quay.io/vessl-ai/torch:2.3.1-cuda12.1
run:
  - workdir: /root
    command: nvidia-smi
4

Add metadata

Finally, let’s polish up by giving our Run a name and description. Here’s the completed quickstart.yaml:

name: Quickstart
description: Run "nvidia-smi"
resources:
  cluster: vessl-oci-sanjose
  preset: gpu-a10-small
image: quay.io/vessl-ai/torch:2.3.1-cuda12.1
run:
  - workdir: /root
    command: nvidia-smi

​
Running the workload

Now that we have a completed YAML, we can run the workload with vessl run.

vessl run create -f quickstart.yaml

Running the command will verify your YAML and show you the current status of the workload. Click the output link in your terminal to see full details and a realtime logs of the Run on the web, including the result of the run command.

​
Behind the scenes

When you vessl run, VESSL AI performs the following as defined in quickstart.yaml:

  1. Launch an empty Python container on the cloud with 1 NVIDIA A10 GPU.
  2. Configure runtime with CUDA compute-capable PyTorch 22.09.
  3. Execute nvidia-smi and print the result.

​
Using our web interface

You can repeat the same process on the web. Head over to your Organization, select a project, and create a New run.

​
What’s next?

Now that you’ve run a barebone workload, continue with our guide to launch a Jupyter server and host a web app.

GPU-accelerated notebook

Launch a Jupyter Notebook server with an SSH connection

Stable Diffusion Playground

Launch an interactive web application for Stable Diffusion

Llama 3.1 Fine-tuing

Fine-tune Llama 3.1-8B with instruction datasets

Llama 3.1 deployment

Serve & deploy vLLM-accelerated Llama 3.1-8B