arXiv Machine Learning By Philip Zmushko, Egor Petrov, Nursultan Abdullaev, Mikhail Khrushchev, Samuel Horv\'ath

One-Step Gradient Delay is Not a Barrier for Large-Scale Asynchronous Pipeline Parallel LLM Pretraining

Read the original on arXiv Machine Learning →

arXiv:2606. 30634v1 Announce Type: new Abstract: Modern large-scale LLM pretraining benefits from utilizing Pipeline Parallelism; however, synchronous implementations leave GPUs idle during pipeline bubbles, wasting computational resources.

Summary generated by The Flow from the publisher's feed. The full article lives at arXiv Machine Learning.

Hugging Face Trending Papers
Jul 27

Certified Parallel-in-Time Sinkhorn for Dynamic Entropic Optimal Transport

Dynamic applications, including optimal-transport Flow Matching, repeatedly solve related entropic optimal transport problems, yet conventional distributed Sinkhorn processes frames sequentially and synchronizes after every iteration. We present TemporalSinkhorn, a parallel-in-time executor that batches future candidates and their repairs without making output accuracy speculative.