FFTrainer: Fast Failover in Large-Language Model Training with Almost-Free State Management

TL;DR

FFTrainer is a system that enhances large language model training by enabling rapid failover and efficient state management, significantly reducing recovery time and GPU utilization loss during node failures.

Contribution

It introduces a novel approach that leverages surplus network capacity for fast state saving and loading, improving robustness and efficiency in LLM training.

Findings

Recovery time reduced by up to 98%

GPU utilization loss mitigated by up to 68%

No significant impact on normal training performance

Abstract

Recent developments in large language models (LLMs) have introduced new requirements for efficient and robust training. As LLM clusters scale, node failures, lengthy recoveries, and bulky checkpoints erode efficiency. Infrequent asynchronous checkpoints trigger costly rollbacks, yet higher frequencies add prohibitive overhead. To address these challenges, we propose FFTrainer, a system designed for robust LLM training. FFTrainer leverages surplus network capacity to quickly save and load states, thereby preventing rollbacks and accelerating recovery. Compared with prior checkpointing approaches, FFTrainer reduces recovery time by up to 98% and mitigates GPU utilization loss by up to 68% without hindering normal training.