Remote GPU Cluster Dashboard

A browser-based dashboard for monitoring remote GPU clusters over SSH. Supports two cluster types — RunAI (Salk, password auth) and DSMLP (UCSD, SSH key + Duo 2FA) — connectable independently or simultaneously. View GPU/CPU/RAM metrics, running processes, run terminal commands, and launch Claude Code sessions from a single browser window.

Features

• Dual-mode login — Dropdown selector to connect RunAI, DSMLP, or both; only connected clusters appear in the dashboard
• Overview — All connected clusters at a glance with GPU utilization, memory, CPU load, RAM, and disk usage
• GPU Detail — Per-GPU stats from nvidia-smi: utilization, memory, temperature, power, and running processes
• Process Viewer — Sortable, filterable process table (ps aux) per cluster
• Interactive Terminal — Tabbed terminal emulator (xterm.js) with one tab per connected cluster, auto-cd to project directory and screen -d -r to attach sessions
• Claude Code — Dedicated tab to launch Claude Code on any connected cluster with a file explorer sidebar and markdown viewer
• File Explorer — Browse remote project files with rendered markdown previews
• Dark / Light Mode — VS Code-inspired themes, saved across sessions

Architecture

RunAI:  browser → FastAPI → SSH (password) → cluster
DSMLP:  browser → FastAPI → SSH (key + Duo) → jump box → kubectl exec → pod
                                               jump box → kubesh → pod (terminal)

• Backend: Python FastAPI with paramiko for SSH
• Frontend: Vanilla HTML/CSS/JS with xterm.js and marked.js loaded from CDN (no npm, no build step)
• Auth: RunAI uses SSH password (held in memory only); DSMLP uses SSH key + Duo 2FA to a jump box

Setup

conda env create -f environment.yml
conda activate gpu-dashboard

Usage

uvicorn app:app --reload --host 0.0.0.0 --port 8000

Open http://localhost:8000. Use the dropdown to select RunAI, DSMLP, or Both, then click Connect. For DSMLP, approve the Duo push on your phone. The dashboard shows only the clusters you connected to.

Configuration

All settings live in config.yaml:

RunAI (Salk)

project:
  directory: /home/jovyan/vast/kaiwen/track-mjx   # file explorer root, terminal auto-cd
  screen_session: train-vqvae                      # auto-attach in Terminal tab (screen -d -r)
  claude_screen_session: claude                    # screen session for Claude tab
  claude_user: devuser                             # su to this user for Claude

clusters:
  my-cluster:
    host: 10.0.0.1
    port: 22
    username: root

• project.directory — file explorer root, terminal auto-cd, and Claude tab working directory
• project.screen_session — auto-attached via screen -d -r in Terminal tab if running
• clusters — SSH targets (host, port, username), connected via password entered at login

DSMLP (UCSD)

DSMLP uses SSH key + Duo 2FA to a jump box (dsmlp-login.ucsd.edu), then manages Kubernetes pods for GPU access.

dsmlp:
  host: dsmlp-login.ucsd.edu
  port: 22
  username: kbian
  key_path: ~/.ssh/id_ed25519                      # SSH key for jump box
  launch_command: "launch-scipy-ml.sh -W DSC180A_FA25_A00 -g 4 -p low -c 8 -m 64 -n 31 -b"
  project:
    directory: private/MOSAIC                      # relative to home, or absolute path
    screen_session: train-vqvae
    claude_screen_session: claude
    claude_user: devuser
  auth_guide: "https://github.com/KevinBian107/DSC180-Toolbox"

• key_path — path to SSH private key (Ed25519)
• launch_command — command to launch a new pod if none is running
• dsmlp.project — same fields as RunAI project, but for the DSMLP pod environment
• directory — can be relative (e.g. private/MOSAIC) or absolute

SSH key setup: See the DSC180 Toolbox guide for generating and registering your SSH key with DSMLP.

Pod lifecycle:

1. Connect — SSH key auth + Duo 2FA (auto-sends push) to the jump box
2. Auto-detect — Checks kubectl get pods for an existing Running pod
3. Launch — If no pod exists, shows "Launch Pod" button; runs the configured launch_command and polls until Running
4. Access — Commands run via kubectl exec, terminals via kubesh

Project Structure

config.yaml         # All configuration (clusters, project, server, dsmlp)
config.py           # Loads config.yaml into Python (exports SERVER, PROJECT, CLUSTERS, DSMLP)
app.py              # FastAPI app — REST + WebSocket endpoints for both RunAI and DSMLP
ssh_manager.py      # SSH connection pool: RunAI (password) + DSMLP (key/Duo/kubectl)
environment.yml     # Conda environment (fastapi, uvicorn, paramiko, pyyaml)
static/
  index.html        # Single-page dashboard UI with dropdown login
  style.css         # VS Code-inspired dark/light theme styles
  app.js            # Frontend: dual-mode login, metrics, terminals, file explorer
docs/
  vqvae.md          # VQVAE training quick reference