DeepSeek Open Source Week has reached its fourth day, where three repositories have been open-sourced. The first one is DualPipe, a bidirectional pipeline parallel algorithm that focuses on optimizing the computational and communication efficiency of large model training. Its core lies in overlapping the forward and backward computation-communication stages, reducing the 'bubble' time in traditional pipeline parallelism. The second repository is EPLB, an expert parallel load balancer that addresses the issue of GPU load imbalance in expert parallelism, ensuring resource utilization. The third repository is a computation-communication overlap analysis, used to observe performance bottlenecks and optimization effects in DeepSeek-V3 training and inference. These open-source initiatives aim to promote collaboration and innovation in the field of artificial intelligence and machine learning.