Skupper + vLLM Deployment Postmortem

Knowledge bundle capturing lessons learned from deploying IBM Granite models on multi-GPU cloud instances via Skupper V2 Virtual Application Networks, August 2026.

Concepts

• Skupper Platform Migration - Why custom-built skrouterd on linux/systemd failed and how the official container image on podman platform solved it; 13+ build iterations; octets=0 mystery; podman gotchas
• Network Architecture - Edge vs interior mode limitation (source code proof); AWS port constraints; routing key design; multi-hub connectivity; link token gotchas; CLI status reporting bug
• vLLM Model Deployment - 128K context VRAM calculations; Granite 4.1 30B/8B tensor parallelism; InstructLab container as vLLM runtime; model selection mistakes (Qwen3 40K, Granite “14B”); cold start times
• AgentFS Process Lessons - Custom provider JSON schema incident; three-layer defensive fix; backup guardrail; skill-index signals bug; context window degradation in long sessions
• Skills Inventory - Complete inventory of skills updated/renamed; AgentFS guardrail changes; final model registry and Skupper VAN architecture