Piper: programmable distributed training system decouples strategy from runtime

Piper, developed by a team including Megan Frisella and Gilbert Bernstein from UC Berkeley and others, tackles a core challenge in large‑scale model training: the tight coupling between high‑level parallelism strategies (data, pipeline, expert, ZeRO) and their low‑level runtime implementations. Current systems either rely on human experts to manually craft execution plans or use general‑purpose frameworks that are locked to a fixed set of strategies. Piper introduces an intermediate representation (IR) — a unified global training DAG that represents all computation and communication. Users declare their strategy through simple model annotations and scheduling directives; each directive transforms the IR. From that IR, the system automatically compiles per‑device execution plans and runs them on a strategy‑agnostic distributed runtime.

This design allows Piper to maintain performance parity with widely used strategies like ZeRO while also enabling novel, optimized schedulers such as DeepSeek‑V3’s DualPipe. DualPipe jointly schedules compute and communication, reducing idle time and memory overhead — something that would be extremely difficult to implement in traditional frameworks. By decoupling strategy from execution, Piper gives researchers and engineers a programmable way to experiment with cutting‑edge parallelism without rewriting low‑level code. The system is detailed in a paper on arXiv (2606.11169) and represents a step toward more flexible, maintainable distributed training infrastructure for foundation models.