DeepSeek Elastic Compute (DSec)
4 hours ago
- Large-scale agentic training with LLMs requires isolated, stateful execution environments for tasks like inspecting repositories, invoking tools, and running commands.
- DSec provides a unified SDK supporting multiple sandbox backends (FnCall, container, microVM, full-VM) to meet diverse isolation and functionality needs.
- It coordinates placement and lifecycle management across clusters, composes environments from independent layers, and uses memory sharing, reclamation, and CPU scheduling for high-density execution.
- DSec loads image data on demand from a cluster-wide distributed filesystem (3FS) and is co-designed with the RL framework to decouple rollout execution from GPU training.
- The platform preserves rollout state while reclaiming idle resources and mitigates agent misbehavior such as reward hacking.
- A single production unit (~160 nodes) serves about 3 million sandboxes daily, supporting over 380,000 concurrent sandboxes and sustaining over 5,000 creations per second.
- Evaluation shows reduced environment setup and image-distribution overhead, improved memory efficiency, and maintained latency-sensitive performance under high-density overcommit.