← sources

Ways to reduce frequent compaction on larger projects?

source: reddit (reddit:LocalLLaMA)published 2026-10-14 15:37ingested 2026-10-15 02:42

retained text

I'm using Qwen 3.8 27B on an R9700 and it works great, occasionally I'll switch to Tiel-coder for rapid simple tasks. One big problem I run into that makes real projects come to a halt is that more code takes up more context, compaction becomes more and more frequent between any edits, until it gets stuck in a compaction loop not doing any useful work. I started using Q4\_K\_XL and Q8 kv cache, for a context of 200k tokens, which helps from my normal 100-150k, but is there a technique to ensure it works efficiently? 

One idea I have but never fully enforced is having the code indexed (function signatures with comments and line numbers, structs shown in full so it doesn't need a read, maybe methods listed under structs.) Or another way is to make the code extremely modular so the LLM will have a very narrow focus. I tried this but it becomes very cumbersome because other modules will also need changed, probably until each becomes mature enough?

document fetched 2026-10-15 00:46 · canonical https://reddit.com/r/LocalLLaMA/comments/1x5uhme/ways_to_reduce_frequent_compaction_on_larger