Efficiently training large language models (LLMs) on commodity cloud resources remains challenging due to limitations in network bandwidth and accelerator memory capacity. Existing training systems ...