LCSB: Layer-Cyclic Selective Backpropagation for Memory-Efficient On-Device LLM Fine-Tuning

TL;DR

LCSB introduces a selective backpropagation method that reduces memory and computation for on-device LLM fine-tuning, maintaining accuracy and improving stability, especially in quantized models.

Contribution

LCSB is a novel approach that computes gradients for only a subset of layers, justified by theoretical analysis as Block Coordinate Descent, enabling efficient on-device LLM fine-tuning.

Findings

Achieves up to 1.40× speedup with minimal quality loss.

Maintains stability and convergence in 4-bit quantized models.

Provides theoretical convergence guarantees for selective backpropagation.

Abstract

Memory-efficient backpropagation (MeBP) has enabled first-order fine-tuning of large language models (LLMs) on mobile devices with less than 1GB memory. However, MeBP requires backward computation through all transformer layers at every step, where weight decompression alone accounts for 32--42% of backward time. We propose Layer-Cyclic Selective Backpropagation (LCSB), which computes gradients for only a subset of layers per step. Our key insight is that residual connections guarantee gradient flow through identity paths, while AdamW momentum provides implicit updates for non-selected layers. We interpret LCSB as Block Coordinate Descent on the LoRA parameter space, providing theoretical justification for convergence. LCSB achieves up to 1.40$\times$ speedup with less than 2\% quality degradation across five models and three tasks. Surprisingly, in 4-bit quantized settings, LCSB exhibits superior stability: a 3B model that completely diverges under full backpropagation converges smoothly with LCSB, suggesting an implicit regularization effect from selective gradient computation.