This story was originally published on HackerNoon at:
https://hackernoon.com/why-local-llms-suddenly-slow-down-at-long-context.
Your local LLM runs fine until it doesn't. A look at KV cache spilling from VRAM into shared memory, and why it happens silently on Windows.
Check more stories related to tech-stories at:
https://hackernoon.com/c/tech-stories.
You can also check exclusive content about
#local-llms,
#llama.cpp,
#kv-cache,
#vram,
#gpu,
#local-inference,
#machine-learning,
#hackernoon-top-story, and more.
This story was written by:
@speederx. Learn more about this writer by checking
@speederx's about page,
and for more stories, please visit
hackernoon.com.
Your local LLM runs fine until the context fills up past a certain point - then generation speed can drop by ~50%. The cause is the KV cache spilling out of VRAM into slower shared memory. On Windows it happens silently, with no out-of-memory error to warn you.