5 papers · 1 filter
One-Step Gradient Delay is Not a Barrier for Large-Scale Asynchronous Pipeline Parallel LLM Pretraining
Philip Zmushko, Egor Petrov, Nursultan Abdullaev +2
Modern large-scale LLM pretraining benefits from utilizing Pipeline Parallelism; however, synchronous implementations leave GPUs idle during pipeline bubbles, wasting computational…
Preconditioned Norms: A Unified Framework for Steepest Descent, Quasi-Newton and Adaptive Methods
Andrey Veprikov, Arman Bolatov, Aleksandr Bogdanov +4
Optimization lies at the core of modern deep learning, yet existing methods often face a fundamental trade-off between adapting to problem geometry and leveraging curvature utiliza…
MT-DAO: Multi-Timescale Distributed Adaptive Optimizers with Local Updates
Alex Iacob, Andrej Jovanovic, Mher Safaryan +6
Training large models with distributed data parallelism (DDP) requires frequent communication of gradients across workers, which can saturate bandwidth. Infrequent communication st…
Differentially Private Clipped-SGD: High-Probability Convergence with Arbitrary Clipping Level
Saleh Vatan Khah, Savelii Chezhegov, Shahrokh Farahmand +2
Gradient clipping is a fundamental tool in Deep Learning, improving the high-probability convergence of stochastic first-order methods like SGD, AdaGrad, and Adam under heavy-taile…
DES-LOC: Desynced Low Communication Adaptive Optimizers for Training Foundation Models
Alex Iacob, Lorenzo Sani, Mher Safaryan +8
Scaling foundation model training with Distributed Data Parallel (DDP) methods is bandwidth-limited. Existing infrequent communication methods like Local SGD were designed to synch…