Avoiding wasteful electricity use while self hosting LLMs
SMRTR summary
When an AI model running on a home server got stuck for 20 hours at 85 watts with no active work, it triggered a search for a permanent fix. The culprit was a known Ollama bug leaving the GPU pegged at ~89% busy. The solution: a lightweight watchdog script that checks GPU load every 5 minutes and sends a phone alert after 15 minutes of suspicious activity — without auto-restarting, to avoid killing real work.
SMRTR provides this summary for quick context. The original article belongs to Hacker News.
Read the original article