But if the LLM stays in the V-RAM or even just stays in normal RAM does it still benefit from zram? I thought that only helped when the ram was not enough.

So I’d say it is more likely the bore scheduler + better drivers.

Basically it compresses your data in the RAM. Needs a little more work form the CPU but it is still faster than swap. fyi

I had zram setup on my previous OS as well and on cachy the LLM didn’t need to use it while on my old OS it did. My guess would be that the driver had a little less overhead.

I did some Benchmarks and CachyOS claims of around 15% more performance seem to be true. Unigin Heavenbenchmark , Super Tuxkart and Furmark all got improved scores compared to PopOS. Additionally Fallout 4 now runs a lot smoother which is probably due to the BORE scheduler doing something better. My local LLMs also seem to be slightly faster and for some reason now need less V-Ram.