Show HN: UATC – A Closed-Loop Controller to Prevent GPU OOM
UATC is a closed-loop controller designed to prevent GPU out-of-memory (OOM) errors by monitoring GPU memory usage and automatically cleaning up unused resources or adjusting allocation.
Background
- UATC (pronounced "U-A-T-C") is a command-line tool that monitors GPU memory in real time and automatically pauses (or slows down) a running process when it's about to hit an out‑of‑memory (OOM) error, then resumes it when memory frees up. It acts as a closed‑loop controller rather than a simple watchdog.
- GPU OOM errors are a common frustration for developers, researchers, and hobbyists running large AI models (e.g., Llama, Stable Diffusion) on a single GPU. When a process exceeds available VRAM, it crashes abruptly, often wasting hours of computation.
- Existing solutions include manual monitoring (nvidia-smi), static memory allocation hacks, or simply buying a more expensive GPU. UATC tries to fill the gap by adding dynamic, feedback‑based control — something that has been missing from free, open‑source tooling.
- The project is hosted on GitHub and was posted to Hacker News as "Show HN", meaning the author is introducing their own creation to the community for feedback.