Dancing along Battery: Enabling Transformer with Run-time Reconfigurability on Mobile Devices

TL;DR

This paper introduces RT3, a framework that combines hardware and software reconfiguration techniques to enable efficient, energy-saving Transformer model execution on mobile devices with real-time switching capabilities.

Contribution

RT3 is the first to integrate hybrid block-structured and pattern pruning with run-time reconfigurability for Transformer models on mobile devices.

Findings

RT3 achieves over 4x battery life extension with less than 1% accuracy loss.

RT3 switches pattern sets within 45ms to meet real-time constraints.

The framework effectively supports dynamic hardware conditions on resource-constrained devices.

Abstract

A pruning-based AutoML framework for run-time reconfigurability, namely RT3, is proposed in this work. This enables Transformer-based large Natural Language Processing (NLP) models to be efficiently executed on resource-constrained mobile devices and reconfigured (i.e., switching models for dynamic hardware conditions) at run-time. Such reconfigurability is the key to save energy for battery-powered mobile devices, which widely use dynamic voltage and frequency scaling (DVFS) technique for hardware reconfiguration to prolong battery life. In this work, we creatively explore a hybrid block-structured pruning (BP) and pattern pruning (PP) for Transformer-based models and first attempt to combine hardware and software reconfiguration to maximally save energy for battery-powered mobile devices. Specifically, RT3 integrates two-level optimizations: First, it utilizes an efficient BP as the first-step compression for resource-constrained mobile devices; then, RT3 heuristically generates a shrunken search space based on the first level optimization and searches multiple pattern sets with diverse sparsity for PP via reinforcement learning to support lightweight software reconfiguration, which corresponds to available frequency levels of DVFS (i.e., hardware reconfiguration). At run-time, RT3 can switch the lightweight pattern sets within 45ms to guarantee the required real-time constraint at different frequency levels. Results further show that RT3 can prolong battery life over 4x improvement with less than 1% accuracy loss for Transformer and 1.5% score decrease for DistilBERT.