Self-Hosted LLM Cluster (vLLM / TGI) Behind the Gateway
For teams with on-prem GPUs — same OpenAI-compatible alias trick, weights running locally, hybrid cloud + on-prem in one gateway, routing by cost or compliance.
💡
Tip: The cleanest argument for an AI gateway is the hybrid story: cloud models for breadth, on-prem models for sensitive workloads, one base URL for application code.
What this use case covers
- vLLM or Text Generation Inference (TGI) serving open-weights models on local GPUs.
- The same OpenAI-compatible alias trick the rest of the course uses, applied to a self-hosted backend.
- Hybrid routing: cloud + on-prem behind a single gateway, with traffic split by cost, latency, or compliance class.
- GPU autoscaling notes (KEDA / Karpenter) and zero-downtime model swaps.