Skip to content

RUNE Documentation

Guide

lpasquali/rune-docs

END-TO-END GUIDE¶

This guide walks you from a blank machine to a fully operational RUNE deployment running scheduled AI benchmarks against a Kubernetes cluster.

Section 1 — Prerequisites¶

kubectl ≥ 1.27
helm ≥ 3.12
Python ≥ 3.11
rune-bench package
Vast.ai API key
Kubernetes cluster access

Section 2 — Deploy RUNE API Server (Helm)¶

kubectl create namespace rune
helm install rune ./charts/rune \
  --namespace rune \
  --set rune.api.authDisabled=false \
  --set rune.api.tokens="myteam:$(openssl rand -hex 24)"

Section 3 — Configure `rune.yaml`¶

Example configuration for your local environment:

version: "1"
defaults:
  model: llama3.1:8b
  question: "What is unhealthy?"
  backend: http
  ollama_warmup: true

Section 4 — Run Your First Benchmark¶

Cloud GPU (Vast.ai)¶

rune --profile cloud-gpu run-benchmark \
  --question "Why is the cluster degraded?" \
  --model llama3.1:8b \
  --vastai-stop-instance

Existing Server¶

rune run-benchmark \
  --ollama-url http://your-ollama-server:11434 \
  --model llama3.1:8b \
  --question "What is unhealthy?"

Section 5 — Cost Philosophy¶

RUNE surfaces spend before provisioning. - Vast.ai: Based on max_dph. - Cloud Stubs: AWS ($2.50/hr), GCP ($2.20/hr), Azure ($3.06/hr). - Local: Based on TDP, energy rates, and amortization.