DeepSeek 开源 DeepEP:首个面向 MoE 模型的 EP 通信库
70
MoE 模型的通信瓶颈一直是训练和推理的痛点,DeepEP 专为此优化,做大规模分布式训练的团队值得关注。
DeepSeek 开源了 DeepEP,这是首个专为 MoE(混合专家)模型设计的专家并行(EP)通信库。它提供了高吞吐、低延迟的 GPU 内核,支持训练和推理中的全到全通信。
DeepEP 还支持低精度操作,如 FP8,并引入了高效的稀疏通信技术。该库已开源在 GitHub 上,开发者可以访问其 Pull Requests 页面了解更多。
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include DeepSeek, MoE, EP通信库, 开源/仓库.
Primary source verification: published directly by DeepSeek: GitHub 新仓库 with first-party authority.