Efficiently Distilling LLMs for Edge Applications
This work addresses the need for cost-effective and scalable model distillation in industrial edge computing, though it appears incremental as it builds on existing supernet training approaches.
The paper tackles the problem of efficiently training supernets for large language models (LLMs) to produce smaller models for edge applications, proposing a method that achieves high-quality encoder models and reduces decoder training time through slicing.
Supernet training of LLMs is of great interest in industrial applications as it confers the ability to produce a palette of smaller models at constant cost, regardless of the number of models (of different size / latency) produced. We propose a new method called Multistage Low-rank Fine-tuning of Super-transformers (MLFS) for parameter-efficient supernet training. We show that it is possible to obtain high-quality encoder models that are suitable for commercial edge applications, and that while decoder-only models are resistant to a comparable degree of compression, decoders can be effectively sliced for a significant reduction in training time.