sites: Inventar und Zielbild an echte Knoten (pve44–pve61) angepasst; ADR-0001/0002/0006 ergänzt
ci / test (push) Successful in 25s
ci / test (push) Successful in 25s
Aufbauplan Task 0.4: - inventory.yaml: pve44/46/48/52/61 und GODE-PC-05/06 (Ada, Windows, cpu_roles: false) - proxmox.yaml: eine GPU-VM pro Knoten mit allen Karten (D2), local-lvm/tank (D1 B), 192.168.100.0/24 ohne VLAN, Gast-IPs .220–.239 (ANNAHME bis Task 0.2) - registry.yaml/prometheus.yml: ops-Labels auf gpu-pve44/46/48, Ada ohne pve_node/vmid - tofu: dynamische hostpci, Speicher je Gast; Ansible: pve_gpu_hosts/pve_kiosk, Replikation auf mehrere Ziele - Placement: cpu_roles=false schließt Arbeitsplätze/Kiosk von CPU-Rollen aus - PLAN §1.3/§1.4/§9 und ADR-Nachträge (D1 entschieden, D2/D3 Bestätigung offen) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
+16
-13
@@ -1,6 +1,9 @@
|
||||
# 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).
|
||||
# Stand Aufbauplan Task 0.4: P2000-Slots in den GPU-VMs gpu-pve44/46/48 (Zuordnung ANNAHME bis Benchmark-Tag,
|
||||
# Task 3.3). Ada-Slots laufen auf Windows-Arbeitsplätzen (D3) ohne Proxmox; Umstellung auf die
|
||||
# Worker-Registry v2 folgt in Task 6.4, bis dahin behalten sie ihre Engine.
|
||||
slots:
|
||||
tier0-strong:
|
||||
tier: 0
|
||||
@@ -15,37 +18,37 @@ slots:
|
||||
tier: 1
|
||||
device: gpu
|
||||
engine: vllm
|
||||
node: gpu-main # RTX 4000 Ada #1
|
||||
node: gpu-main # RTX 4000 Ada, GODE-PC-05 (Windows, D3)
|
||||
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
|
||||
ops: {docker_host: gpu-main, container: vllm} # kein Proxmox (D3); bis Task 6.4
|
||||
|
||||
tier1-flex:
|
||||
tier: 1
|
||||
device: gpu
|
||||
engine: vllm
|
||||
node: gpu-flex # RTX 4000 Ada #2 (nachts Training → unavailable)
|
||||
node: gpu-flex # RTX 4000 Ada, GODE-PC-06 (Windows, D3)
|
||||
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
|
||||
ops: {docker_host: gpu-flex, container: vllm} # kein Proxmox (D3); bis Task 6.4
|
||||
|
||||
verifier-small:
|
||||
tier: 1
|
||||
device: gpu
|
||||
engine: llamacpp
|
||||
node: gpu-small # Quadro M4000 (Nutzen per Benchmark prüfen)
|
||||
node: gpu-small # Quadro P2000 in gpu-pve48
|
||||
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
|
||||
ops: {docker_host: gpu-pve48, container: llamacpp-chat, pve_node: pve48, vmid: "248"} # ANNAHME bis Task 3.3
|
||||
|
||||
signature-extractor:
|
||||
tier: 1
|
||||
@@ -56,18 +59,18 @@ slots:
|
||||
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
|
||||
ops: {docker_host: gpu-pve48, container: llamacpp-chat, pve_node: pve48, vmid: "248"} # ANNAHME bis Task 3.3
|
||||
|
||||
embed-default:
|
||||
tier: 1
|
||||
device: gpu
|
||||
engine: llamacpp
|
||||
node: gpu-embed # Quadro P2000
|
||||
node: gpu-embed # Quadro P2000, redundant in gpu-pve44 und gpu-pve46
|
||||
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
|
||||
ops: {docker_host: gpu-pve44, container: llamacpp-embed, pve_node: pve44, vmid: "244"} # ANNAHME bis Task 3.3
|
||||
|
||||
rerank-default:
|
||||
tier: 1
|
||||
@@ -78,7 +81,7 @@ slots:
|
||||
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
|
||||
ops: {docker_host: gpu-pve44, container: llamacpp-rerank, pve_node: pve44, vmid: "244"} # ANNAHME bis Task 3.3
|
||||
|
||||
# --- Sprachausgabe (ADR-0006) ---
|
||||
tts-de:
|
||||
@@ -90,17 +93,17 @@ slots:
|
||||
base_model: piper-de-thorsten-medium
|
||||
max_context: 5000 # Zeichen pro Anfrage (TTS-Dienst-Limit)
|
||||
kind: tts
|
||||
ops: {docker_host: infra, container: tts-piper, pve_node: pve1, vmid: "101"} # ANNAHME bis Deployment
|
||||
ops: {docker_host: infra, container: tts-piper, pve_node: pve44, vmid: "101"} # HA: kann nach pve52/pve48 wandern
|
||||
|
||||
tts-de-hq:
|
||||
tier: 1
|
||||
device: gpu
|
||||
engine: pytorch
|
||||
node: gpu-flex # teilt sich die Ada mit tier1-flex; während Training unavailable
|
||||
node: gpu-flex # teilt sich die Ada mit tier1-flex; unter Windows erst ab Phase 9 (WSL2, D3)
|
||||
litellm_model: tts-de-hq
|
||||
base_model: qwen3-tts-1.7b
|
||||
max_context: 5000
|
||||
kind: tts
|
||||
ops: {docker_host: gpu-flex, container: qwen-tts, pve_node: pve-ada, vmid: "202"} # ANNAHME bis Deployment
|
||||
ops: {docker_host: gpu-flex, container: qwen-tts} # kein Proxmox (D3)
|
||||
|
||||
adapters: []
|
||||
|
||||
Reference in New Issue
Block a user