An auto-parallel training method for deep learning models with extracting model structural features

Yan Zeng

doi:10.1016/j.asoc.2026.115426

An auto-parallel training method for deep learning models with extracting model structural features

2026-05-14

0

PRE

AI

DOI：10.1016/j.asoc.2026.115426

Original

Original request for help

Save

Abstract

En 中文

• We construct a performance evaluation model for model parallel technology, balancing communication, computation, and memory loads to eliminate resource bottlenecks. • We use the Q-Actor-Critic algorithm to improve sampling efficiency and accelerate the convergence of reinforcement learning. • We combine evolution strategy and Q-Actor-Critic to search for optimal parallel strategies, leveraging global optimization and self-learning in dynamic environments. • Experiments show ESRL reduces search time by 22.4% and per-step execution time by 24.7% vs. Placeto, with 5.8% better performance while balancing resources.

Keywords:

model parallelism

reinforcement learning

evolution strategy

performance optimization

deep learning

An auto-parallel training method for deep learning models with extracting model structural features

Abstract

Journal

Organization