llama-server Kept Crashing Overnight on a 6GB Laptop: System RAM, Not the Model

Stunning-Sherbet1853 · reddit · 2026-10-14

While building a Windows local agent on Ollama (llama.cpp), the author ran the same overnight task kit on a 24GB desktop, 8GB laptop, and 6GB laptop — the 6GB machine crashed three times in one night. Two culprits:

1. llama.cpp's prompt cache

llama-server keeps prompt cache in system RAM, not VRAM; default limit is 8GB, and one 40K-token state is 1.3GB
On a 24GB-RAM laptop the runner grew to 12GB; the author now sets --cache-ram relative to system RAM (4GB on the 24GB machine)

2. Their own agent process leaking

18GB leaked after 12 hours on the 8GB laptop: network responses kept after browser tabs closed, plus a Windows component definition that only matched the dev PC
Found by counting live objects by type and diffing two snapshots 10 minutes apart

Before/after (same task, same model)

8GB laptop: 3,280→5,434MB in 101 min before; 2,699→2,809MB in 91 min after
6GB laptop: zero crashes over 5 hours after fixes (runner leveled at 9GB)

Takeaway: when a local LLM looks unstable on a small GPU, check who's using system RAM before changing model size or quantization — VRAM estimates don't include any of this.

Original post →

llama-server Kept Crashing Overnight on a 6GB Laptop: System RAM, Not the Model

More from coding & agent