SinoTechIntel Academic Portal
Official PDF TranslationFrontiers of Information Technology & Electronic Engineering

Automatic parallelism strategy generation with minimal memory redundancy

Authors: Yanqi SHI; Peng LIANG; Hao ZHENG; Linbo QIAO; Dongsheng LI

DOI: 10.1631/FITEE_2300684Status: Verified Translated Edition
Sponsored AdvertisementAd Placement Area
reCAPTCHA Bot Shield Active

Preparing Secure Academic Download

Verifying human reader & generating high-resolution document...

Verifying Document Integrity15s remaining
← Back to Article
Protected by Google reCAPTCHA v3.PrivacyTerms
Sponsored ContentAdSense In-Feed Ad Slot

Key Findings in This Report

• Proposes a novel algorithm for automatic parallelism strategy generation with minimal memory redundancy. • Introduces a redundant memory cost model to compute memory overhead per operator in distributed training. • Formulates the parallelism strategy search as an integer linear programming (ILP) problem for optimal intra-operator parallelism. • Achieves up to 67% memory savings versus Megatron-LM while maintaining comparable throughput.