Zum Inhalt springen
TakoAPI
Zurück zu den Skills

homelab-cluster

von mlesnews

DevOps & Cloud
609Downloadzahlen laut clawskills.sh
2Sterne
0Aufrufe
0 LikesMelde dich an, um diesen Skill zu liken
—Noch nicht bewertet
Deine Bewertung

Kurzbeschreibung

Manage multi-tier AI inference clusters for homelabs.

Beschreibung

Was dieser Skill macht

Manages multi-tier AI inference clusters for homelab environments. Covers health monitoring across GPU and CPU nodes, MoE-based request routing, automatic recovery for downed nodes, and model deployment via Ollama and llama.cpp. Includes GPU memory planning, Docker volume strategies, and a unified LiteLLM API gateway.

Consolidates health checks, failover routing, and node recovery into one workflow instead of managing each Ollama or llama.cpp instance separately.

Wann du ihn einsetzt

  • Routing coding tasks to a dedicated coder model on the local GPU node
  • Recovering a remote Docker Desktop crash without manual intervention
  • Planning which quantized models fit within 24GB VRAM across a cluster
  • Restarting GPU containers sequentially to avoid CUDA initialization deadlocks
  • Exposing all cluster models through a single OpenAI-compatible endpoint

Beispiel-Workflow

EINGABE

User asks: route this code review request to the best available model

AGENT
  1. 1

    Classify the prompt as a 'code' task

  2. 2

    Query the health endpoint for the designated coder model node

  3. 3

    If healthy, return the coder model endpoint; if degraded, check fallback tier

  4. 4

    Select next-best available model from health results

  5. 5

    Return the chosen model endpoint and tier metadata to the caller

AUSGABE

Request routed to remote/mark-i running Qwen2.5-Coder-7B at full GPU capacity

Voraussetzungen

  • SSH access credentials stored in a vault (e.g., Azure Key Vault or HashiCorp Vault) for remote node recovery

Installation

clawhub install mlesnews/homelab-cluster

Führe den CLI-Befehl aus oder füge den Prompt in Claude Code, Cursor, Codex oder deinen Coding-Agenten ein.

Links