This project is the high-performance inference framework used for DeepSeek's large-scale online services, providing core capabilities for offline inference and large-scale deployment. It integrates self-developed optimized operators such as FlashMLA, DeepGEMM, and DeepEP, and supports attention data parallelism, MoE expert parallelism, asynchronous scheduling, deterministic sampling, and structured output. It can run on NVIDIA and Huawei Ascend NPUs.