万字解析DeepSeek MOE架构！

Core Viewpoint - The article provides a comprehensive overview of the Mixture of Experts (MoE) architecture, particularly focusing on the evolution and implementation of DeepSeek's MoE models (V1, V2, V3) and their optimizations in handling token distribution and load balancing in AI models [2][21][36]. Group 1: MoE Architecture Overview - MoE, or Mixture of Experts, is a model architecture that utilizes multiple expert networks to enhance performance, particularly in sparse settings suitable for cloud computing [2][3]. - The initial interest in MoE architecture surged with the release of Mistral.AI's Mixtral model, which highlighted the potential of sparse architectures in AI [2][3]. - The Switch Transformer model introduced a routing mechanism that allows tokens to select the top-K experts, optimizing the processing of diverse knowledge [6][10]. Group 2: DeepSeek V1 Innovations - DeepSeek V1 addresses two main issues in existing MoE practices: knowledge mixing and redundancy, which hinder expert specialization [22][24]. - The model introduces fine-grained expert division and shared experts to enhance specialization and reduce redundancy, allowing for more efficient knowledge capture [25][26]. - The architecture includes a load balancing mechanism to ensure even distribution of tokens across experts, mitigating training inefficiencies [32]. Group 3: DeepSeek V2 Enhancements - DeepSeek V2 builds on V1's design, implementing three key optimizations focused on load balancing [36]. - The model limits the number of devices used for routing experts to reduce communication overhead, enhancing efficiency during training and inference [37]. - A new communication load balancing loss function is introduced to ensure equitable token distribution across devices, further optimizing performance [38]. Group 4: DeepSeek V3 Developments - DeepSeek V3 introduces changes in the MOE layer computation, replacing the softmax function with a sigmoid function to improve computational efficiency [44]. - The model eliminates auxiliary load balancing losses, instead using a learnable bias term to control routing, which enhances load balancing during training [46]. - A sequence-level auxiliary loss is added to prevent extreme imbalances within individual sequences, ensuring a more stable training process [49].