Strix Halo LLM Serving: 25 tok/s at 151k Context Under 100W26 April 2026·7 minsAi Homelab Llm-Inference Strix-Halo Llama-Cpp Qwen Local-Ai