DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale
Summary
DSec presents a production sandbox platform with multiple backend sandboxes and a unified SDK to support large-scale agentic training. It coordinates placement, layering, memory sharing, and on-demand data loading, while addressing safety concerns like reward hacking. The paper reports production-scale metrics, including 160-node units serving millions of sandboxes per day and high concurrency, with improvements in setup time and memory efficiency.