spark-pack

Pack models onto 1–3 NVIDIA DGX Sparks before you start them. Each machine is a 121 GB tank. The verdict is FITS, TIGHT, or won't start.

At a glance

What it is	A visual memory packer for DGX Spark unified memory — measured residents, mutual exclusion, live probe.
What it’s for	So you do not hard-reboot a box by co-residing H3 with DS4F, or treating 2×121 GB as one pool.
How to use it	Open index.html, or ./setup.sh → http://127.0.0.1:8768/

Try it (pick one)

One command

git clone https://github.com/Coinupbtc/spark-pack.git
cd spark-pack && ./setup.sh

Copy-paste

git clone https://github.com/Coinupbtc/spark-pack.git && cd spark-pack
python3 -m http.server 8768 --bind 127.0.0.1
# open http://127.0.0.1:8768/

No Spark

Open index.html in a browser. The catalog is the product. ./probe.sh is optional and only reads free, ports, and docker names.

What the tanks mean

121 GB is per Spark, not a shared cluster pool. TP2 DS4F puts ~79 GB of weights on each rank. That is why two Sparks look full, not half empty.

Verdict	Meaning
FITS	≥15 GB headroom on the fullest Spark
TIGHT	Fits, but earlyoom country. Live DS4F clusters live here.
Won't start	Over 121 GB, or H3 on one Spark, or a measured kill combo

H3 + DS4F is the measured kill combo (89.4 GB on Spark 1). Vision 4B can sit on Spark 2 beside the cluster.

docker stats under-reports unified memory. Believe free and the dated sizes in catalog.json.

Live probe

On a Spark:

./probe.sh
# optional second node:
SPARK_PACK_PEER=user@other-spark ./probe.sh

Then refresh the packer and hit Match live. Nothing is started or stopped.

Not this repo

• Starting DS4F — use the upstream DSpark recipe, not this packer.
• Generic “does 70B q4 fit?” math — that is sparkfit.
• A results museum — every catalog row has an as_of date.

License

MIT