Rethinking Memory and Communication Costs for Efficient Data Parallel Training of Large Language Models | Read Paper on Bytez