Presentation

· Contributors · Organizations · Search Program · My Schedule · Happening Now

This content is available for: Technical Program Reg Pass, Workshop Reg Pass. Upgrade Registration

LLMTailor: A Layer-wise Tailoring Tool for Efficient Checkpointing of Large Language Models

SessionPDSW'25: The 10th International Parallel Data Systems Workshop

DescriptionCheckpointing is essential for fault tolerance in training large language models (LLMs). However, existing methods, regardless of I/O strategies, periodically store the entire model and optimizer states, incurring substantial storage overhead and contention. Recent studies reveal that updates across LLM layers are highly non-uniform. During training, some layers may undergo more significant changes, while others remain stable or even unchanged. This suggests that selectively checkpointing only layers with significant updates could reduce overhead without harming training. Implementing such strategies requires fine-grained control over both weights and optimizer states, which no current tool provides. To address this gap, we propose LLMTailor, a checkpoint-merging-framework that filters and assembles layers from different checkpoints to form a composite checkpoint. Our evaluation indicates that LLMTailor can work with different checkpointing strategies and effectively reduce checkpoint size (e.g., 4.3 times smaller for Llama3.1-8B) and checkpoint time (e.g., 2.8 times faster for Qwen2.5-7B) while maintaining model quality.

Author/Presenters