pipelining

This paper introduces techniques to manage memory peaks in training large Mixture-of-Experts models with long context lengths, including Pipelined LLEP, Ring-DTP, SCO, and OffloadStreamAdamW, which enable fixed GPU working sets and improve throughput up to 10.4x.

StreamMA introduces a streaming communication paradigm for multi-agent reasoning that pipelines intermediate results to reduce latency and improve effectiveness by leveraging more reliable early steps, outperforming baselines across benchmarks and revealing a step-level scaling law.