Skip to main content

Manage Pods with dstack on RunPod

dstack is an open-source tool that simplifies the orchestration of Pods for AI and ML workloads. By defining your application and resource requirements in YAML configuration files, it automates the provisioning and management of cloud resources on RunPod, allowing you to focus on your application logic rather than the infrastructure.

In this guide, we'll walk through setting up dstack with RunPod to deploy vLLM. We'll serve the meta-llama/Llama-3.1-8B-Instruct model from Hugging Face using a Python environment.

Prerequisites​

  • A RunPod account with an API key
  • On your local machine:
    • Python 3.8 or higher
    • pip (or pip3 on macOS)
    • Basic utilities: curl
  • These instructions are applicable for macOS, Linux, and Windows systems.
note

Windows Users

  • It's recommended to use WSL (Windows Subsystem for Linux) or tools like Git Bash to follow along with the Unix-like commands used in this tutorial
  • Alternatively, Windows users can use PowerShell or Command Prompt and adjust commands accordingly

Installation​

Setting Up the dstack Server​

  1. Prepare Your Workspace

    Open a terminal or command prompt and create a new directory for this tutorial:

    mkdir runpod-dstack-tutorial
    cd runpod-dstack-tutorial

  2. Set Up a Python Virtual Environment

    python3 -m venv .venv
    source .venv/bin/activate

  3. Install dstack

    Use pip to install dstack:

    pip3 install -U "dstack[all]"

    Note: If pip3 is not available, you may need to install it or use pip.

Configuring dstack for RunPod​

  1. Create the Global Configuration File

    The following config.yml file is a global configuration used by dstack for all deployments on your computer. It's essential to place it in the correct configuration directory.

    • Create the configuration directory:

      mkdir -p ~/.dstack/server

    • Navigate to the configuration directory:

      cd ~/.dstack/server

    • Create the config.yml File

      In the configuration directory, create a file named config.yml with the following content:

      projects:
        - name: main
          backends:
            - type: runpod
              creds:
                type: api_key
                api_key: YOUR_RUNPOD_API_KEY

      Replace YOUR_RUNPOD_API_KEY with the API key you obtained from RunPod.

  2. Start the dstack Server

    From the configuration directory, start the dstack server:

    dstack server

    You should see output indicating that the server is running:

[INFO] Applying ~/.dstack/server/config.yml...
[INFO] The admin token is ADMIN-TOKEN
[INFO] The dstack server is running at http://127.0.0.1:3000
note

The ADMIN-TOKEN displayed is important for accessing the dstack web UI.

  1. Access the dstack Web UI
  • Open your web browser and navigate to http://127.0.0.1:3000.
  • When prompted for an admin token, enter the ADMIN-TOKEN from the server output.
  • The web UI allows you to monitor and manage your deployments.

dstack web ui

Deploying vLLM as a Task​

Step 1: Configure the Deployment Task​

  1. Prepare for Deployment

  • Open a new terminal or command prompt window.

  • Navigate to your tutorial directory:

    cd runpod-dstack-tutorial

  • Activate the Python Virtual Environment

    source .venv/bin/activate
  1. Create a Directory for the Task

Create and navigate to a new directory for the deployment task:

mkdir task-vllm-llama
cd task-vllm-llama
  1. Create the dstack Configuration File

  • Create the .dstack.yml File

    Create a file named .dstack.yml (or dstack.yml if your system doesn't allow filenames starting with a dot) with the following content:

    type: task
    name: vllm-llama-3.1-8b-instruct
    python: "3.10"
    env:
      - HUGGING_FACE_HUB_TOKEN=YOUR_HUGGING_FACE_HUB_TOKEN
      - MODEL_NAME=meta-llama/Llama-3.1-8B-Instruct
      - MAX_MODEL_LEN=8192
    commands:
      - pip install vllm
      - vllm serve $MODEL_NAME --port 8000 --max-model-len $MAX_MODEL_LEN
    ports:
      - 8000
    spot_policy: on-demand
    resources:
      gpu:
        name: "RTX4090"
        memory: "24GB"
      cpu: 16..
note

Replace YOUR_HUGGING_FACE_HUB_TOKEN with your actual Hugging Face access token (read-access is enough) or define the token in your environment variables. Without this token, the model cannot be downloaded as it is gated.

Step 2: Initialize and Deploy the Task​

  1. Initialize dstack

Run the following command in the directory where your .dstack.yml file is located:

dstack init
  1. Apply the Configuration

Deploy the task by applying the configuration:

dstack apply

  • You will see an output summarizing the deployment configuration and available instances.

  • When prompted:

    Submit the run vllm-llama-3.1-8b-instruct? [y/n]:

    Type y and press Enter to confirm.

  • The ports configuration provides port forwarding from the deployed pod to localhost, allowing you to access the deployed vLLM via localhost:8000.

  1. Monitor the Deployment

  • After executing dstack apply, you'll see all the steps that dstack performs:

    • Provisioning the pod on RunPod.
    • Downloading the Docker image.
    • Installing required packages.
    • Downloading the model from Hugging Face.
    • Starting the vLLM server.

  • The logs of vLLM will be displayed in the terminal.

  • To monitor the logs at any time, run:

    dstack logs vllm-llama-3.1-8b-instruct

  • Wait until you see logs indicating that vLLM is serving the model, such as:

    INFO: Started server process [1]
    INFO: Waiting for application startup.
    INFO: Application startup complete.
    INFO: Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)

Step 3: Test the Model Server​

  1. Access the Service

Since the ports configuration forwards port 8000 from the deployed pod to localhost, you can access the vLLM server via http://localhost:8000.

  1. Test the Service Using curl

Use the following curl command to test the deployed model:

curl -X POST http://localhost:8000/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
          "model": "meta-llama/Llama-3.1-8B-Instruct",
          "messages": [
             {"role": "system", "content": "You are Poddy, a helpful assistant."},
             {"role": "user", "content": "What is your name?"}
          ],
          "temperature": 0,
          "max_tokens": 150
        }'
  1. Verify the Response

You should receive a JSON response similar to the following:

{
  "id": "chat-f0566a5143244d34a0c64c968f03f80c",
  "object": "chat.completion",
  "created": 1727902323,
  "model": "meta-llama/Llama-3.1-8B-Instruct",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "My name is Poddy, and I'm here to assist you with any questions or information you may need.",
        "tool_calls": []
      },
      "logprobs": null,
      "finish_reason": "stop",
      "stop_reason": null
    }
  ],
  "usage": {
    "prompt_tokens": 49,
    "total_tokens": 199,
    "completion_tokens": 150
  },
  "prompt_logprobs": null
}

This confirms that the model is running and responding as expected.

Step 4: Clean Up​

To avoid incurring additional costs, it's important to stop the task when you're finished.

  1. Stop the Task

In the terminal where you ran dstack apply, you can stop the task by pressing Ctrl + C.

You'll be prompted:

Stop the run vllm-llama-3.1-8b-instruct before detaching? [y/n]:

Type y and press Enter to confirm stopping the task.

  1. Terminate the Instance

The instance will terminate automatically after stopping the task.

If you wish to ensure the instance is terminated immediately, you can run:

dstack stop vllm-llama-3.1-8b-instruct
  1. Verify Termination

Check your RunPod dashboard or the dstack web UI to ensure that the instance has been terminated.

Additional Tips: Using Volumes for Persistent Storage​

If you need to retain data between runs or cache models to reduce startup times, you can use volumes.

Creating a Volume​

Create a separate dstack file named volume.dstack.yml with the following content:

type: volume
name: llama31-volume

backend: runpod
region: EUR-IS-1

# Required size
size: 100GB
note

The region ties your volume to a specific region, which then also ties your Pod to that same region.

Apply the volume configuration:

dstack apply -f volume.dstack.yml

This will create the volume named llama31-volume.

Using the Volume in Your Task​

Modify your .dstack.yml file to include the volume:

volumes:
- name: llama31-volume
 path: /data

This configuration will mount the volumne to the /data directory inside your container.

By doing this, you can store models and data persistently, which can be especially useful for large models that take time to download.

For more information on using volumes with RunPod, refer to the dstack blog on volumes.

Conclusion​

By leveraging dstack on RunPod, you can efficiently deploy and manage Pods, accelerating your development workflow and reducing operational overhead.