core: Inventar/Placement, Konfig-Validierung, CLI, HTTP-API, LiteLLM-Konfiguration, Beispiel-Skills
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
@@ -0,0 +1,51 @@
|
||||
# Modell-Katalog (SPEC-inventory §2). vram_gb = Gewichte + Mindest-KV-Cache (Richtwerte, per Kalibrierung prüfen).
|
||||
# Ist ein Modell nicht mehr aktuell, wird ein neuer Eintrag ergänzt. Der Wechsel läuft über die Lifecycle-Pipeline.
|
||||
models:
|
||||
qwen3-14b-awq:
|
||||
kind: llm
|
||||
vram_gb: 14
|
||||
min_cc: 7.5 # AWQ-Kernels in vLLM ab Turing
|
||||
engines: [vllm]
|
||||
context: 32768
|
||||
lora: true
|
||||
qwen3-8b-awq:
|
||||
kind: llm
|
||||
vram_gb: 9
|
||||
min_cc: 7.5
|
||||
engines: [vllm]
|
||||
context: 32768
|
||||
lora: true
|
||||
qwen3-8b-q4km:
|
||||
kind: llm
|
||||
vram_gb: 7
|
||||
min_cc: 5.0
|
||||
engines: [llamacpp]
|
||||
context: 16384
|
||||
qwen3-4b-q4km:
|
||||
kind: llm
|
||||
vram_gb: 4
|
||||
min_cc: 5.0
|
||||
engines: [llamacpp]
|
||||
context: 8192
|
||||
cpu_ok: true # langsam, aber als Notfall-Verifier nutzbar
|
||||
qwen3-embedding-0.6b-q8:
|
||||
kind: embedding
|
||||
vram_gb: 1.5
|
||||
min_cc: 5.0
|
||||
engines: [llamacpp]
|
||||
context: 8192
|
||||
cpu_ok: true
|
||||
bge-m3-f16:
|
||||
kind: embedding
|
||||
vram_gb: 2.5
|
||||
min_cc: 5.0
|
||||
engines: [llamacpp]
|
||||
context: 8192
|
||||
cpu_ok: true
|
||||
bge-reranker-v2-m3-q8:
|
||||
kind: reranker
|
||||
vram_gb: 1.5
|
||||
min_cc: 5.0
|
||||
engines: [llamacpp]
|
||||
context: 8192
|
||||
cpu_ok: true
|
||||
@@ -0,0 +1,34 @@
|
||||
# Rollen (SPEC-inventory §2/§3). Placement arbeitet sie nach priority ab.
|
||||
roles:
|
||||
gpu-main:
|
||||
priority: 1
|
||||
needs: {gpu: true, engine: vllm, min_vram_gb: 16, min_cc: 7.5}
|
||||
prefer: largest
|
||||
kinds: [llm]
|
||||
slots: [tier1-general]
|
||||
gpu-flex:
|
||||
priority: 2
|
||||
needs: {gpu: true, engine: vllm, min_vram_gb: 16, min_cc: 7.5}
|
||||
prefer: largest
|
||||
kinds: [llm]
|
||||
optional: true # ohne zweite große GPU: kein Fallback-Slot, Training nur nachts auf gpu-main
|
||||
slots: [tier1-flex]
|
||||
gpu-embed:
|
||||
priority: 3
|
||||
count: 2 # redundant, damit Routing einen GPU-Ausfall übersteht
|
||||
needs: {gpu: true, engine: llamacpp, min_vram_gb: 3, min_cc: 5.0}
|
||||
prefer: smallest_fitting
|
||||
kinds: [embedding, reranker]
|
||||
cpu_fallback: true
|
||||
slots: [embed-default, rerank-default]
|
||||
gpu-small:
|
||||
priority: 4
|
||||
needs: {gpu: true, engine: llamacpp, min_vram_gb: 4, min_cc: 5.0}
|
||||
prefer: smallest_fitting
|
||||
kinds: [llm]
|
||||
cpu_fallback: true
|
||||
slots: [verifier-small, signature-extractor]
|
||||
infra:
|
||||
priority: 5
|
||||
needs: {gpu: false, min_ram_gb: 16}
|
||||
slots: []
|
||||
Reference in New Issue
Block a user