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

TL;DR

LLMTailor, a checkpoint-merging framework that filters and assembles layers from different checkpoints to form a composite checkpoint, is proposed, which can work with different selective checkpointing strategies and effectively reduce checkpoint size and checkpoint time while maintaining model quality.

Abstract

Checkpointing is essential for fault tolerance in training large language models (LLMs). However, existing methods, regardless of their I/O strategies, periodically store the entire model and optimizer states, incurring substantial storage overhead and resource contention. Recent studies reveal that updates across LLM layers are highly non-uniform. Across training steps, some layers may undergo more significant changes, while others remain relatively stable or even unchanged. This suggests that selectively checkpointing only layers with significant updates could reduce overhead without harming training. Implementing such selective strategies requires fine-grained control over both weights and optimizer states, which no current tool provides. To address this gap, we propose \texttt{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 selective 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.

Figures (3)

• Figure 1: The layer-wise structure in a Llama3.1-8B model.
• Figure 2: The AdamW optimizer used in a Llama3.1-8B model.
• Figure 3: Reconstruct the parameter groups in the optimizer before training.