Browser terminal

The browser terminal is how you get a shell inside a running GPU instance. It runs over a WebSocket from the console, so nothing to install locally.

Open the terminal

1. Go to Compute → GPU Instances in the console.
2. Click into the running instance you want to access.
3. Click the Terminal button in the top-right of the detail page.

A full-screen terminal opens in your browser. You're root inside the container — standard shell behavior, every command you'd expect (ls, cd, python, vim, apt, nvidia-smi, etc.) works.

Behavior and limits

• Single active session per instance — opening the terminal from a second tab closes the first. Use tmux or screen inside if you need multiple panes.
• Your browser tab is the connection — if the tab closes or you lose network, the WebSocket drops. Your process inside the container keeps running (it's not tied to the tab); just reopen the terminal to reattach.
• Copy / paste works with native browser shortcuts (Cmd/Ctrl+C to copy the selection, Cmd/Ctrl+V to paste).
• Idle timeout — connections idle for > 60 min are closed server-side. Reopen to reconnect.

Long-running jobs — use tmux

The terminal is your window into the container, not the container itself. A training job running in python train.py keeps running if you close the tab — but you lose visibility unless you reattach to a tmux session:

# First time:
apt-get update && apt-get install -y tmux
tmux new -s work

# Now inside tmux — run your long job:
python train.py

# Detach without stopping: Ctrl-b then d
# Close the tab. Come back tomorrow. Open the terminal:
tmux attach -t work

Best practice: run any job expected to take more than a few minutes inside tmux.

Browser terminal vs Jupyter

Prefer the terminal for	Prefer Jupyter for
Installing packages, system setup	Interactive data exploration
Running scripts, long training jobs	Plotting, visualizations
Debugging shell / environment issues	Iterative code you want to re-run incrementally
git, vim, htop, nvidia-smi	Notebook-based ML workflows

Both are available on the same instance; use whichever fits the task.

Troubleshooting

"Disconnected from server" shortly after opening

• Your instance's network might be flaky. Refresh the detail page; if the instance shows status preempted or failed, the pod went down and the terminal can't attach.

Keyboard shortcuts aren't working

• Some browser extensions intercept key combos. Try an incognito window.

Command output looks garbled

• The terminal emulator is VT100-compatible. Programs that auto-detect terminal type via $TERM should work; if not, export TERM=xterm-256color before running.

GPU isn't visible

• Run nvidia-smi to confirm. If it says "No devices found," the pod might not have been scheduled onto a GPU node — very rare, but try terminating and relaunching.

Common quick-checks inside the instance

nvidia-smi                         # GPU status
df -h                              # disk free
free -h                            # RAM
nproc                              # CPUs available
python -c "import torch; print(torch.cuda.is_available(), torch.cuda.device_count())"