llama.cpp flag tuning boosted generation 70% on 40GB VRAM eGPU laptop

Benchmarking llama.cpp flags on a 40GB VRAM laptop with TB4 eGPU raised generation speed from 16 to 27 t/s (+70%), prefill from 376 to 573, and usable context to the full 262k. User also filed a bug in llama around MTP; final command included Qwen3.8-27B-UD-Q6_K_XL.gguf with -c 262144.