Best settings for Qwen 3.6 27B on 48GB VRAM setup

User shares optimal llama.cpp settings for running Qwen 3.6 27B (Q8_0) across RTX 4090 + RTX 3090 using tensor split. Discussion covers VRAM management and inference tuning on the dual-GPU system.