core: Inventar/Placement, Konfig-Validierung, CLI, HTTP-API, LiteLLM-Konfiguration, Beispiel-Skills

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-12 00:47:13 +02:00
parent f2da893f71
commit 4cbe313c62
20 changed files with 1085 additions and 19 deletions
@@ -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: []