← AI Terminology

Model Serving

Model serving is the production practice of hosting models behind APIs with batching, scaling, routing, observability, and SLOs — not just running generate() in a notebook.

It is the MLOps/LLMOps discipline of reliable inference.
Why It Matters in AI
Research demos die without serving: autoscaling, multi-model routing, canaries, auth, and cost control. LLM serving adds KV cache, continuous batching, and streaming semantics on top of classic model servers.
Key Points
Aspect Description
Obs TTFT, TPS, queue depth, GPU util, errors
Stacks vLLM, TGI, TensorRT-LLM, Triton, SageMaker, KServe
Related LLMOps, continuous batching
Concerns Latency, throughput, cost, reliability, security
Patterns Blue/green, canary, A/B, shadow traffic
LLM specifics Streaming, prefix cache, LoRA multiplex
Simple Analogy
A restaurant kitchen during dinner rush — not one home meal, but stations, tickets, timing, and quality under load.
Common Usage Examples
  • Deploy vLLM OpenAI-compatible server
  • Kubernetes GPU deployments
  • Canary 5% traffic to new weights
  • Dashboards for GPU cache and latency
Summary
In short: Model serving is production hosting of models with scale, SLOs, and ops — the difference between a demo and a product.