Quartet: Native FP4 Training Can Be Optimal for Large Language Models | Read Paper on Bytez