35 lines
1.1 KiB
YAML
35 lines
1.1 KiB
YAML
# 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: []
|