Spark Console

Local GPU + fleet dashboard for NVIDIA DGX Spark (Grace Blackwell, unified memory) — and any Linux box with an NVIDIA GPU. Pulse, fleet cards, a read-only services board. Nothing phones home.

Default UI: http://127.0.0.1:8085/

This public tree is the portable core: collectors, the UI, and a read-only service board. It runs standalone. The hero screenshot is a live dual-Spark run on the author’s box (your clone shows your hosts and units, not that fleet).

Published default is main. Ignore leftover master if GitHub still lists it — that branch is not this dashboard.

At a glance

What it is	Spark Console — a local GPU/system dashboard for NVIDIA DGX Spark: live graphs, model inventory, endpoint health, and a read-only service board.
What it’s for	See GPU, memory, model and service health in one page on localhost — without phoning home, and without a cloud account.
How to use it	./setup.sh → open http://127.0.0.1:8085/. Stop with ./stop.sh.

Try it (pick one)

One command

git clone https://github.com/Coinupbtc/spark-console.git
cd spark-console && ./setup.sh
# open http://127.0.0.1:8085/

Copy-paste

git clone https://github.com/Coinupbtc/spark-console.git && cd spark-console
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
./start.sh

Already cloned

./setup.sh

Runtime data stays in local data/ (gitignored). Nothing phones home.

What you get

Path	What
console.html	The console UI
server.py	Read-only HTTP API (see the route list in its docstring)
collector.py	Performance snapshots → data/ (CSV + latest JSON)
services_lite.py	Service board driven by your services.json
remote_node.py	Optional second node over SSH
cluster_metrics.py	Two-node metrics + degraded-node alerting
baseline_summary.py	Percentiles, availability, longest-run report
todos_lite.py	Local scratch list for the console panel
setup.sh / start.sh / stop.sh	One-command lifecycle

The service board

Copy the example and edit it — the board is empty until you do:

cp services.example.json services.json

Each entry needs an id plus at least one of unit, port, or probe_url:

{"services": [
  {"id": "ollama", "label": "Ollama", "group": "inference",
   "unit": "ollama.service", "port": 11434, "critical": true}
]}

critical: true raises an alert when that service is not running. The board is read-only — the console reports state and never starts or stops anything.

Optional second node

Point it at another machine over SSH and its metrics join the graphs:

export [email protected]     # or NODE2_SSH_ALIAS=mynode from ~/.ssh/config
export NODE2_SSH_KEY=~/.ssh/id_ed25519   # pin an identity (needed under systemd)
./start.sh

Unset, the console is simply a single-node dashboard — no errors, no alerts.

Environment

Var	Default	Meaning
PORT / HOST	8085 / 127.0.0.1	Bind address
CONSOLE_ALLOWED_HOSTS	—	Extra Host values to accept (see Security)
NODE2_HOST / NODE2_SSH_ALIAS	—	Enable the second node
NODE2_SSH_KEY	—	SSH identity for the second node
NODE2_ENDPOINTS	8000,8080,11434	Remote ports probed for /v1/models
SPARK_CONSOLE_SERVICES	./services.json	Service board config path
SPARK_CONSOLE_SYSTEMD_SCOPE	--user	Use --system for system units
NOTIFY_HOOK	—	Executable taking --plain <msg>, paged when node2 degrades
SPARK_FABRIC_IP	—	If set, also probe :8800 on that address (no baked-in LAN IP)
DGX_DATA_DIR	./data	Where snapshots and the CSV live

Security

The console binds 127.0.0.1 only, which is necessary but not sufficient — a localhost HTTP server is still reachable from any page your browser loads. Two defences are built in:

• Host allowlist — blocks DNS rebinding, where a page on evil.com re-points its own hostname at 127.0.0.1 to become same-origin with the console and read every endpoint. Foreign Host headers get a 400.
• Origin check on writes — a cross-origin POST with no custom headers is a "simple request", so no CORS preflight would stop it. Foreign Origin values get a 403. Requests with no Origin (curl, scripts) still work.

To reach the console from another machine, prefer an SSH tunnel — the Host stays localhost, so nothing needs configuring:

ssh -L 8085:127.0.0.1:8085 user@your-spark

If you must serve a real hostname, add it explicitly:

CONSOLE_ALLOWED_HOSTS=spark.lan ./start.sh

This tree has no start/stop, restart, model-switch, kill, or launch HTTP routes. Pulse, fleet cards, and the service board are read-only. POST /api/todos is a local scratch list; POST /api/refresh re-runs the collector.

Tests

./.venv/bin/python -m pytest tests/ -q

Related

• miaai35-tune
• spark-training-lab
• zwell-bench

License

MIT