For Strix Halo - Official llama.cpp isn't ideal and how to highest possible throughput

Official llama.cpp is not optimized for Strix Halo (gfx1151), achieving less than 50% of its hardware's theoretical performance. For optimal throughput, users should consider peonist-ai/halogen-flash-server, which is specifically optimized for Strix Halo and the Qwen 3.8 Flash Next (Q38FN) model. This setup, also known as "Ninfer for Strix Halo," can reach approximately 50 tokens/second for decoding and 1200 tokens/second for prefill, utilizing about 90% of the hardware's theoretical capacity.