# Registry für den Cluster "knaster" (SPEC-core §1).
# Clients sprechen nur Slot-Namen an; litellm_model ist der Alias in deploy/litellm/config.yaml.
slots:
  tier0-strong:
    tier: 0
    device: external
    engine: openrouter
    litellm_model: tier0-strong
    base_model: openrouter-strong       # konkretes Modell steht nur in der LiteLLM-Konfiguration
    max_context: 200000
    effective_context: 120000

  tier1-general:
    tier: 1
    device: gpu
    engine: vllm
    node: gpu-main                       # RTX 4000 Ada #1
    litellm_model: tier1-general
    base_model: qwen3-14b-awq            # Startkandidat, Entscheidung per Kalibrierungslauf
    max_context: 32768
    effective_context: 16384
    supports_lora: true
    fallbacks: [tier1-flex]

  tier1-flex:
    tier: 1
    device: gpu
    engine: vllm
    node: gpu-flex                       # RTX 4000 Ada #2 (nachts Training → unavailable)
    litellm_model: tier1-flex
    base_model: qwen3-14b-awq
    max_context: 32768
    effective_context: 16384
    supports_lora: true

  verifier-small:
    tier: 1
    device: gpu
    engine: llamacpp
    node: gpu-small                      # Quadro M4000 (Nutzen per Benchmark prüfen)
    litellm_model: verifier-small
    base_model: qwen3-4b-q4km
    max_context: 8192
    effective_context: 6144

  signature-extractor:
    tier: 1
    device: gpu
    engine: llamacpp
    node: gpu-small
    litellm_model: signature-extractor
    base_model: qwen3-4b-q4km
    max_context: 8192
    effective_context: 6144

  embed-default:
    tier: 1
    device: gpu
    engine: llamacpp
    node: gpu-embed                      # Quadro P2000
    litellm_model: embed-default
    base_model: qwen3-embedding-0.6b-q8
    max_context: 8192
    kind: embedding

  rerank-default:
    tier: 1
    device: gpu
    engine: llamacpp
    node: gpu-embed
    litellm_model: rerank-default
    base_model: bge-reranker-v2-m3-q8
    max_context: 8192
    kind: rerank

adapters: []