vLLM deployment

Hands-on project to run vLLM locally, then deploy it on Kubernetes — including bootstrapping the cluster from greenfield code.

This repo is the source of truth for phase constraints, Loop 2 decisions, validation ladders, and check scripts. Obsidian (secondbrain/coding/1-projects/vllm-deployment/Deploy vLLM Locally and on K8s.md) is the project index only.

Project phases

Phase	Goal	Status	Doc
1 — Local	vLLM-Metal on Apple Silicon	Done	phase1/README.md
2 — Cluster	kubeadm cluster on AWS (EICE, kubectl on node)	Done	phase2/README.md
3 — WireGuard	VPN access — laptop kubectl replaces EICE	Planned	phase3/README.md
4 — vLLM deploy	vLLM CPU; in-cluster API works	Planned	phase4/README.md
5 — GPU	NVIDIA/CUDA	Planned	phase5/README.md
6 — Operate	Expose externally; metrics	Planned	phase6/README.md

Phase 1        Phase 2           Phase 3          Phase 4         Phase 5        Phase 6
(local)        (cluster)         (WireGuard)      (vLLM)          (GPU)          (operate)

Reference architecture: ~/DEV/k8s-homelab — greenfield code in this repo.

Repository layout

vllm-deployment/
├── README.md
├── phase1/          # local vLLM-Metal
├── phase2/          # cluster factory (TF, Packer, kubeadm)
├── phase3/          # WireGuard access
├── phase4/          # vLLM workload (Helm, phase4-check.sh)
├── phase5/          # GPU extension (planning)
└── phase6/          # expose & observe (planning)

Design notes

• Phase 2 ops: kubectl on node only (homelab smoke-test style). Validated: make check-full and post-reboot make check-cluster (2026-08-23).
• Phase 3: WireGuard unlocks laptop-native Helm/kubectl for Phase 4+.
• Validation-driven: each phase completes when its check script passes.