DeepEP is an open-source communication library developed by the Chinese AI company DeepSeek. It is designed to accelerate the all-to-all communication used in expert parallelism, where a Mixture of Experts (MoE) model is spread across multiple GPUs and tokens must be dispatched to the GPUs holding each expert and gathered back. It was released in February 2025 during DeepSeek's open-source week.
It supports both NVLink within a server and RDMA between servers, and provides separate high-throughput kernels for training and prefill and low-latency kernels for decoding. It is one of the libraries on which DeepSeek's inference engine DeepInfer is built.