Deploying a large language model (LLM) spans several artifacts and lifecycle steps.
The container image, model checkpoint, infrastructure configuration, and running graphics processing unit (GPU) memory are related, but they are not the same thing. Keeping them separate makes startup failures and rollout trade-offs much easier to explain.
In this chapter
- What actually gets deployedSeparate software images, model artifacts, configuration, and runtime state.
- Where model weights liveFollow a checkpoint from registry or object storage to disk, RAM, and GPU video memory (VRAM).
- From container start to ready replicaTrace scheduling, image pull, model loading, warmup, and readiness.
- Updating models without breaking trafficReason about rolling updates, canaries, surge capacity, and rollback when replicas are expensive to start.
For the mechanics of images, Pods, Deployments, Services, and probes, use Infrastructure & DevOps. These pages focus on what changes once the workload is an LLM.