DeepSeek has declared this week as Open Source Week and plans to consecutively open source five software libraries. Recently, the company announced the public release of DeepEP. Within just 20 minutes of the announcement, DeepEP garnered over 1,000 stars on Microsoft's (MSFT, Financial) GitHub, a code hosting platform.
DeepEP is an expert parallelism communication library designed for the training and inference of mixture of experts (MoE) models. It facilitates efficient and optimized all-to-all communication, supporting low-precision computations including FP8, which meets modern high-performance computing demands. Moreover, DeepEP has been deeply optimized for asymmetric bandwidth forwarding scenarios from NVLink to RDMA. It not only provides high throughput but also supports control over the number of Streaming Multiprocessors, balancing both training and inference tasks for high throughput performance.
