Building a Disciplined Local AI Workstation: VRAM Gating and Lifecycle Management

How do you run heavy Multimodal LLMs, VLMs, and Whisper models concurrently on a single 16GB GPU without OOM crashes?

In our open-source project GoodQ4All, we built a Python-based ModelLifecycleManager context manager that audits system VRAM via PyTorch and nvidia-smi, performs preflight checks against strict budget profiles, and automatically unloads resident models.

Here is the step-by-step architecture: https://github.com/GoodQ02/goodq4all