RIFT: A Scalable Methodology for LLM Accelerator Fault Assessment using Reinforcement Learning

Khurram Khalil, Muhammad Mahad Khaliq, Khaza Anuarul Hoque

The massive scale of modern AI accelerators presents critical challenges to traditional fault assessment methodologies, which face prohibitive computational costs and provide poor…