Switch llama-server invocations from --parallel to -np with -kvu (kv-cache unified) across Qwen3.6 model configs. Also reduce context for qwen3.6-27b-cuda0 from 150k to 140k.
Switch llama-server invocations from --parallel to -np with -kvu (kv-cache unified) across Qwen3.6 model configs. Also reduce context for qwen3.6-27b-cuda0 from 150k to 140k.