DeepSeek open source DeepEP – library for MoE training and Inference

DeepEP is a communication library tailored for Mixture-of-Experts (MoE) and expert parallelism (EP), providing high-throughput and low-latency GPU kernels for efficient AI parallel processing tasks like MoE dispatch and combine.

• Efficient expert-parallel library
• Supports high-throughput GPUS
• Tailored for MoE and EP
• Offers low-latency operations
• Compatible with Hopper GPUs