Ditching llm-scaler: Qwen3.8-27B on my B70 Updated 2026-09-12 Leaving llm-scaler behind, patching vLLM with GPT-5.6 Sol, and getting MTP4 working. My setup as of August 20.