SSH Port Forwarding Workflow for GPU Development

Before reaching for Tailscale or WireGuard, check whether simple SSH port forwarding solves your problem. For an individual developer accessing services on a dedicated GPU server, it usually does.

Contents

• Local forwarding
• Persistent tunnels
• Multiple services
• When to move beyond

Local Forwarding

To access vLLM on port 8000 of the GPU server locally:

ssh -L 8000:localhost:8000 user@gpu-server

While the SSH session is open, localhost:8000 on your laptop proxies to localhost:8000 on the server. Point your browser or client there.

Persistent

For always-on forwarding use autossh:

sudo apt install autossh
autossh -M 0 -N -L 8000:localhost:8000 user@gpu-server

Or set up via systemd user unit for auto-start on login:

[Unit]
Description=SSH tunnel to GPU
[Service]
ExecStart=/usr/bin/autossh -M 0 -N -L 8000:localhost:8000 user@gpu-server
Restart=always
[Install]
WantedBy=default.target

Multiple Services

ssh -L 8000:localhost:8000 \
    -L 8888:localhost:8888 \
    -L 3000:localhost:3000 \
    user@gpu-server

Forwards vLLM (8000), Jupyter (8888), and Grafana (3000) simultaneously.

Or configure in ~/.ssh/config:

Host gpu
    HostName gpu-server.gigagpu.com
    User yourname
    LocalForward 8000 localhost:8000
    LocalForward 8888 localhost:8888
    LocalForward 3000 localhost:3000

Then just ssh gpu activates all forwards.

When to Move Beyond

SSH forwarding stops scaling when:

• Multiple team members need access – manage SSH keys for each
• Non-SSH clients need to connect (mobile apps)
• You want reverse forwarding (server initiates)

At that point move to Tailscale or WireGuard.

See Tailscale and remote VS Code.