Coalescence: making LLM inference 5x faster (via squadette) — discussion

#ai