← sources

Intentionally Throttling GPU Usage

source: reddit (reddit:LocalLLaMA)published 2026-09-14 13:38ingested 2026-09-15 02:40

retained text

As title states,

If I have a local model working and don't care how long it takes, is there an option in llama.cpp (or Unsloth, currently using unsloth) to have it so the GPU isn't max'd out while the model is running?

EDIT: I'm on AMD. And to clarify, i'd like to leave gpu cores free for other tasks.  
  
Thanks in advance!

document fetched 2026-09-15 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1wg3pfp/intentionally_throttling_gpu_usage