DeepSeek open source second wave: EP communication library is here, expected to further reduce computational consumption

DeepSeek has made numerous innovations to efficiently utilize GPUs, and on the 25th, the second day of "Open Source Week," DeepSeek open-sourced the DeepEP communication library. DeepSeek stated that this is the first open-source EP communication library for MoE (Mixture of Experts) model training and inference, which is expected to further reduce computational consumption. According to a report by Yicai, DeepSeek indicated that DeepEP features efficient and optimized communication among all participants; it supports NVLink and RDMA (Remote Direct Memory Access, a communication technology) both within and between nodes; it includes high-throughput kernels pre-filled for training and inference; low-latency kernels for inference decoding; native FP8 scheduling support; and flexible GPU resource control to achieve overlapping computation and communication. EP stands for expert parallelism, a technique used in large-scale distributed AI model training that enhances model parallel processing capabilities and training efficiency. DeepSeek explained on the code hosting site GitHub that for latency-sensitive inference decoding tasks, DeepEP includes a set of low-latency kernels that use pure RDMA to minimize latency. DeepEP also introduces a method for overlapping communication and computation that does not occupy SM (Streaming Multiprocessor) resources. In short, DeepEP is also one of the key technologies for improving GPU utilization efficiency. The performance of DeepSeek-R1, comparable to OpenAI's o1, is based on the model trained with DeepSeek-V3, which is known for its efficient use of advanced NVIDIA GPUs and low training budgets without large-scale usage. To train large models on existing GPUs, DeepSeek has made numerous innovations to efficiently utilize GPU computing power. When asked online about the impact of the open-source DeepEP communication library, DeepSeek responded that DeepEP can significantly enhance the training and inference efficiency of MoE models, greatly reduce computational resource consumption, and that open-sourcing DeepEP helps lower the development costs of AI technology and reduces redundant development. DeepSeek previously announced that it would successively open-source five code repositories this week. Including the code repository FlashMLA open-sourced on February 24, DeepSeek has already open-sourced two code repositories, with three more pending. DeepSeek stated in a previous announcement that it is a small company exploring AGI (Artificial General Intelligence) and, as part of the open-source community, every line of code shared contributes to accelerating the development of the AI industry. ![](https://imageproxy.pbkrs.com/https://pgw.udn.com.tw/gw/photo.php/query-dT1odHRwczovL3VjLnVkbi5jb20udHcvcGhvdG8vMjAyNS8wMi8yNS9yZWFsdGltZS8zMTU1Nzk1Ny5qcGcmeD0wJnk9MCZzdz0wJnNoPTAmc2w9VyZmdz0xMDUwJmV4cD0zNjAwJmV4cD0zNjAw? x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg) DeepSeek has successively open-sourced 5 code repositories. Reuters