AI mania is afoot.  Useful? Yes.  End the world?  Probably not.  Meanwhile there is a lot of bare-metal GPU hardware sitting idle waiting for its next job. In the same spirit as QUADS, I’ve been working on a new project to maximize the usage of GPU systems: QIIP, a tractor beam for spare GPU capacity.

What Is QIIP?
QIIP (QUADS Idle Inference Proxy) pulls free, GPU-equipped bare-metal systems with no current active QUADS schedules and serves them through one OpenAI-compatible inference endpoint. Non-QUADS systems can be manually added through the interface.

Why We Built It
We had the same old problem, just with GPUs instead of servers: fleets of expensive hardware sitting idle between scheduled jobs, with nothing smart to do with it. The goal was a single gateway, user access control, and a token tracker for bring-your-own inference, in the QUADS spirit of squeezing every drop of utilization out of bare-metal.
QIIP only ever reaches for completely free systems and cycles them back out on any new schedule, so the pool grows and shrinks with real demand. It is truly wringing the washcloth of spare capacity.

The Highlights

• One OpenAI-compatible gateway: /v1/chat/completions, /v1/completions, /v1/models with streaming, failover and health checks
• Ultra Efficiency: only free, unscheduled systems join the pool, and they cycle back out on any new schedule
• Engine setup by GPU layout: reads the node GPU model, VRAM and count, then builds and configures vLLM or llama.cpp
• No model guesswork: llmfit and HuggingFace give hardware-aware model picks and ranked recommendations
• User access control: Google OAuth SSO, admin roles, endpoint-scoped qiip_ tokens, per-token usage tracking
• QUADS-native: polls QUADS inventory, unified node view, automatic teardown on scheduling conflicts

Where To Go From Here
Everything is open source at github.com/quadsproject/qiip, with install details and the full feature list in the README. If you want the deeper tour, the project page has the architecture, dashboards and tech stack. If your lab is full of GPUs that only work a few hours a day, this might be the missing piece.