Strata with CachyOS achieves higher throughput than Unsloth on a 9t/s to 48t/s range; 32k context size fits 8GB GPU, 42k feasible but 64k requires more VRAM

I had 9t/s on my rig, with Unsloth 3bit Qwen 3.8 Flash Next. Then I tried STRATA with CachyOS, because I needed something with a ready NVIDIA driver, kinda pain getting the latest NVIDIA driver to work. Cachy just...