Infrastructure Stack

Production-grade orchestration patterns for local LLMs.

RUNTIME
Docker
Standard OCI
ORCHESTRATION
K8s
Scalability Layer
INGRESS
Nginx
Reverse Proxy
OBSERVABILITY
Grafana
Metrics & Logs
> INFRA_PRIMER

The GPU is only the engine — production local AI is an infrastructure stack: an inference engine executes the model (llama.cpp, MLX), a runtime makes it usable (Ollama), a serving layer makes it scale (vLLM/TGI, whose paged attention + continuous batching multiply a GPU's throughput 10–20× under concurrency), and around them sit the chat UI (Open WebUI), the vector DB for RAG (Chroma, Qdrant, pgvector) and the ops layer — Docker with GPU passthrough, pinned versions, Prometheus metrics.

The three reference shapes: solo box (Ollama + one GPU, zero ceremony), team server (vLLM + Open WebUI + SSO on a 48–80GB card behind Docker Compose), production fleet (vLLM instances behind a load balancer, k8s, autoscaling, small-model routing). The classic failure is serving a team on a single-user runtime: it queues, and users blame the model.

Deep references: the full software stack · reference architectures · private ChatGPT build. Below: the latest infrastructure intelligence from the radar.

> SYSTEM_OFFLINE

No infrastructure articles found.

< AI-RADAR MAIN