One RTX 4090 lab server
The lab deployment uses one 24 GB RTX 4090 on a dedicated Ubuntu 22.04/24.04 amd64 host. Ansible prepares the host, K3s and Argo CD. Internal Forgejo/GitLab stores the lab's configuration. A local updater follows signed successful VideoQ releases through Argo CD. The Git server and Argo management API do not need public internet ingress.
For a first installation, use the Japanese step-by-step lab manual, including Mac preparation, internal Forgejo, acceptance tests, updates and rental shutdown.
Use the lab installation runbook and automatic update runbook. They cover SSH/LAN settings, repository keys, initial administrator setup, internal CA trust, backups, explicit initial acceptance and unattended application updates. The lab stores adopted application images locally before deployment; no write access or deployment-registration PR to the application GitHub repository is required.
The supported GPU profile runs GPT-OSS 20B, EmbeddingGemma 2 text-only 768d and Whisper large-v3-turbo in separate GPU-sharing Deployments, using unchanged digest-pinned upstream vLLM images. Audio extras are pinned and provisioned separately. No own engine image is built.
Recorded application and recovery acceptance covers real models, actual UI upload/cited playback and a four-volume cold restore. Recorded automatic-update acceptance covers a real timer-driven release transition through Forgejo and Argo CD, retained videos, and four-volume recovery on a rented RTX 4090. Lab hardware, load testing and actual process-crash recovery remain outside that acceptance.
The former standalone RTX 4090 shell/Textual installer and preliminary Runpod probe have been retired. Use the Ansible/GitOps path for new lab hosts. The generic single-server and CPU development tools remain separate supported paths.
For rental tests, keep the GPU running only during active real-host tests. Stop it during code/CI waits and after verification; stopped-disk storage can still incur charges.