So, llama.cpp has the -nkvo ( --no-kv-offload ) option to offload KV cache to RAM instead of VRAM. Many people avoid this because obviously it hurts performance. But every option exists with a trade off. And in my case, I think it's worth it. Hear me out. I'm running Qwen3.6 27B (IQ4_XS) on RTX 5060 Ti 16GB and 32GB DDR5. In order to fit 65k context, I have to quantize the KV cache down to q4_0, and keep only 58 layers on the GPU. This gives me 23 tps at peak, down to 16 tps during long g...