| name | ray-start |
| description | Start the Ray cluster across all 4 DGX Spark nodes. Copies the per-node start-ray scripts, starts spark-01 as head, then joins spark-02/03/04 with 10-second pauses between each. Verifies cluster health with ray status afterward. |
Ray Cluster Start
Start the Ray cluster on all 4 DGX Spark nodes in the correct order with proper pacing.
What It Does
- Copy scripts — Copies each
cluster/start-ray-spark-*.sh to $HOME/models/start-ray.sh on the corresponding node.
- Start spark-01 (head) — Runs
start-ray.sh from the vllm directory (where .venv lives) on spark-01.
- Wait 10 seconds.
- Start spark-02 — Same pattern, via SSH.
- Wait 10 seconds. Repeat for spark-03, then spark-04.
- Verify — Runs
ray status on spark-01 to confirm all nodes are active.
Usage
./scripts/start_cluster.sh
Run from the skill directory, or from anywhere — the script locates the repo root and env.sh automatically.
Stopping the Cluster
CRITICAL — Stop in reverse order: spark-04 → spark-03 → spark-02 → spark-01. Workers must be stopped before the head node.
./scripts/stop_cluster.sh
Prerequisites
- vLLM installed at
$HOME/models/vllm/.venv on all nodes.
- Passwordless SSH to all Sparks.
env.sh at repo root with SPARK_01_IP through SPARK_04_IP defined.cluster/start-ray-spark-{a,b,c,d}.sh scripts present in the repo.
Before Reinstalling vLLM
CRITICAL — Always stop the Ray cluster before running clean_vllm.sh or reinstalling vLLM. Wiping the venv while Ray is running will leave the cluster in a broken state.
