GPUForge V1 — Demo Walkthrough (10-Step Reference)

Version: 1.0 Audience: Customer engineers running their first GPUForge V1 demo. Scope: End-to-end 10-step demo script extracted from the V1 Reference Design §10.

This walkthrough is the canonical first-pass demo script for GPUForge V1. Follow the ten steps in order — each step is one owner-action or a single artifact reference, so the full demo is a straight read-through of the linked doc artifacts.

1. Cluster setup

Stand up the EKS cluster. Use §2.1 eksctl (preferred) and only fall back to §2.2 AWS CLI when eksctl is unavailable.

Artifact: kubernetes-deployment.md §2 — Cluster Creation

2. Image pull

Install the NVIDIA GPU Operator Helm release (pinned to v26.3.3) and confirm the nvidia.com/gpu resource plus DCGM exporter scrape targets come up.

Artifact: kubernetes-deployment.md §4 — Install GPU Operator

3. First inference job (signal)

Run the CUDA nbody smoke test. Pass signal: non-zero GFLOP/s — the canonical "GPU is wired correctly" indicator.

Artifact: kubernetes-deployment.md §7 — CUDA Smoke Test

4. Multi-tenant path

Wire the multi-tenant gate. Confirm api_keys.key_prefix issuance and tenant-scoped slurm_jobs rows match the E-R diagram, then point tenants at SLURM via the REST API.

Artifact: DEPLOYMENT.md §1 — E-R Diagram and SLURM_SETUP.md §1 — Enable slurmrestd

5. Live-GPU delivery on GKE

Deploy on the g4dn-mng nodegroup configuration referenced in this guide.

Artifact: kubernetes-deployment.md §3.1 — g4dn-mng nodegroup

6. Live-GPU delivery on EKS

Deploy on the p4d-spot nodegroup configuration and confirm spot-interruption handling per §9 pitfall #10.

Artifact: kubernetes-deployment.md §3.2 — p4d-spot nodegroup

7. Live-GPU delivery on-prem

Bring up the on-prem stack via Compose. This is the same launcher as scripts/run.sh used by the Quickstart.

Artifact: QUICKSTART.md — Launch

8. Verify the routing stack

Optional step — only relevant if GAIE + llm-d are installed. Verify EPP status, InferencePool accepted, and a non-zero llm_d_kv_cache_hit_ratio. Skip per §10.6 if inference routing is out of scope.

Artifact: kubernetes-deployment.md §10.5 — Verify the Routing Stack

9. Documented pitfalls

Review the operator-observed pitfalls before the demo. Every operator hits at least one.

Artifact: kubernetes-deployment.md §9 — Pitfalls

10. End-to-end bash recipe

Use the paste-ready bootstrap script to reproduce the cluster end-to-end. Save as eks-cluster-bootstrap.sh, chmod +x, run on a workstation with eksctl + helm + aws + kubectl.

Artifact: kubernetes-deployment.md §11 — End-to-End Bash Recipe