Reproduce “PocketLLM: Enabling On-Device Fine-Tuning for Personalized LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dan Peng, Zhihui Fu, Jun Wang
PocketLLM introduces a method for fine-tuning large language models directly on resource-constrained mobile devices using derivative-free optimization. The approach successfully fine-tuned RoBERTa-large consuming 4GB memory and OPT-1.3B consuming 6.5GB on an OPPO Reno 6 smartphone, demonstrating the feasibility of privacy-preserving personalization.
Key insights
- Memory capacity is the most critical bottleneck preventing LLM fine-tuning on mobile devices, not solely computational speed.
- Derivative-free optimization provides a viable mechanism to drastically reduce memory footprint, thereby enabling on-device fine-tuning on commercial smartphones.
- While feasible, the computational efficiency of LLM fine-tuning on mobile devices remains significantly lower (e.g., 1000x slower) compared to dedicated GPU hardware.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.