Xiaohongshu's Dots Infra team open-sourced BigMac, a pipeline parallel training paradigm for multi-modal LLMs. It nests encoder/generator computation into an optimized LLM schedule, breaking the trade-off between compute speed and memory usage. On understanding tasks, BigMac achieves 1.08-1.9x speedup over baselines with stable memory; on generation tasks, it avoids OOM while delivering 1.5-1.9x speedup.