Service Exposure Guide

Expose network ports from your GPU instances to the internet for APIs, web UIs, and other services.

Overview

Service Exposure allows you to expose network ports from your GPU instances to the internet, making services accessible via public URLs. This eliminates the need for SSH tunneling.

When you expose a service:

Your internal port is mapped to an external port
An external IP address and port are assigned
Traffic from the internet is routed to your pod
You receive a public URL to access the service

Common Use Cases

vLLM API Server

Expose your vLLM OpenAI-compatible API:

Port: 8000
Protocol: TCP

python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port 8000

Jupyter Notebook

Access Jupyter Lab or Notebook remotely:

Port: 8888
Protocol: TCP

TensorBoard

Monitor training metrics in real-time:

Port: 6006
Protocol: TCP

Gradio / Streamlit

Expose custom web applications:

Gradio Port: 7860
Streamlit Port: 8501

How to Expose a Service

Step 1: Start Your Application

Ensure your service is running and listening on the desired port. Important: Always bind to 0.0.0.0 (all interfaces), not localhost or 127.0.0.1.

# Example: Start vLLM API server
python -m vllm.entrypoints.openai.api_server \
  --model your-model \
  --host 0.0.0.0 \
  --port 8000

Step 2: Click "Expose Port"

In the Exposed Services section of your GPU card, click the "+ Expose Port" button.

Step 3: Fill in Service Details

Field	Description	Example
Service Name	A friendly name (lowercase, hyphens allowed)	`vllm-api`
Port	Port your app is listening on	`8000`
Protocol	TCP (default) or UDP	`TCP`

Step 4: Access Your Service

Once exposed, you'll see the external URL:

http://34.123.45.67:31234

Click the copy button to copy the URL.

Managing Services

Viewing Exposed Services

All exposed services are shown in the "Exposed Services" section of your GPU card with their external URLs.

Deleting a Service

Find the service in the Exposed Services section
Click the delete (trash) icon
Confirm the deletion

Note: External clients will lose access immediately.

Best Practices

Security

Add authentication to sensitive services
Use HTTPS when possible for production
Limit exposed ports - only expose what you need
Delete services when no longer needed

Naming

Use descriptive names: vllm-api, jupyter-main
Lowercase only, use hyphens not spaces

Troubleshooting

Service Not Accessible

Application not listening on 0.0.0.0:

# Wrong - only localhost can connect
python app.py --host 127.0.0.1 --port 8000

# Correct - external traffic can connect
python app.py --host 0.0.0.0 --port 8000

Application not running:

ps aux | grep python

Wrong port number:

netstat -tlnp | grep 8000

Connection Timeout

Verify your GPU instance is running
Check the service exists in the Exposed Services section
Ensure your application is running and listening

Alternative: SSH Port Forwarding

For temporary access without exposing publicly, you can use SSH port forwarding instead:

# Forward remote port 8000 to local port 8000
ssh -p <ssh-port> -L 8000:localhost:8000 ubuntu@<host>

Then access at http://localhost:8000 on your machine.

Service Exposure