84 lines
2.8 KiB
YAML
84 lines
2.8 KiB
YAML
# Registry für den Cluster "knaster" (SPEC-core §1).
|
|
# Clients sprechen nur Slot-Namen an; litellm_model ist der Alias in deploy/litellm/config.yaml.
|
|
# ops: Labels für den Healer (Portainer-Endpoint, Container, Proxmox-Node, VM-ID).
|
|
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]
|
|
ops: {docker_host: gpu-main, container: vllm, pve_node: pve-ada, vmid: "201"} # VM-IDs/Hosts: ANNAHME bis Deployment
|
|
|
|
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
|
|
ops: {docker_host: gpu-flex, container: vllm, pve_node: pve-ada, vmid: "202"} # VM-IDs/Hosts: ANNAHME bis Deployment
|
|
|
|
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
|
|
ops: {docker_host: gpu-small, container: llamacpp-chat, pve_node: pve3, vmid: "231"} # VM-IDs/Hosts: ANNAHME bis Deployment
|
|
|
|
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
|
|
ops: {docker_host: gpu-small, container: llamacpp-chat, pve_node: pve3, vmid: "231"} # VM-IDs/Hosts: ANNAHME bis Deployment
|
|
|
|
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
|
|
ops: {docker_host: gpu-embed-1, container: llamacpp-embed, pve_node: pve3, vmid: "221"} # VM-IDs/Hosts: ANNAHME bis Deployment
|
|
|
|
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
|
|
ops: {docker_host: gpu-embed-1, container: llamacpp-rerank, pve_node: pve3, vmid: "221"} # VM-IDs/Hosts: ANNAHME bis Deployment
|
|
|
|
adapters: []
|