软件与服务：AI产业速递：DeepSeek开源DeepEP代码库，优化模型并行处理能力

Investment Rating - The industry investment rating is "Positive" and maintained [6] Core Insights - On February 25, 2025, DeepSeek open-sourced the DeepEP codebase, which is the first open-source expert parallel (EP) communication library for training and inference of mixture of experts (MoE) models. This library allows for distributed training by allocating different experts to various computing devices, leveraging the sparse activation feature of MoE to linearly scale model size with the number of devices without increasing computational costs [2][4] Summary by Sections Event Description - DeepSeek's release of the DeepEP codebase marks a significant advancement in the field of AI, particularly in optimizing model parallel processing capabilities. The library addresses the communication efficiency challenges between experts, significantly reducing data exchange overhead and enhancing training and inference efficiency [4][9] Event Commentary - The DeepEP codebase improves GPU communication efficiency through several methods: 1. Support for NVLink and RDMA within and between GPU nodes, optimizing memory usage without the need for expensive tensor parallelism [9] 2. High-throughput kernels for training and low-latency kernels for inference, enhancing processing speed during both phases [9] 3. Efficient all-to-all communication mechanisms that accelerate information transfer between nodes [9] 4. Flexible GPU resource control that allows for computation-communication overlap, minimizing idle time during training [9] - The report suggests that the new wave of technological supply will lead to a revaluation of the domestic AI industry, enhancing application deployment speed and expanding AI computing demand. Key areas to focus on include: 1. The inference computing power supply chain in China, particularly leading AI chip companies like Cambricon 2. Cloud service providers collaborating with DeepSeek 3. IDC firms working with major companies like Tencent, Alibaba, and ByteDance 4. AI application-related targets in sectors such as government, finance, healthcare, and education [9]