Dynamic control of accuracy and complexity in real time based on available power in micro-LLM

The problem of energy consumption of large language models (LLM) when deployed on battery-powered devices with strict energy constraints is considered. The concept of micro-LLMs is proposed, equipped with mechanisms for dynamically adapting their computational complexity and numerical accuracy in real time, based on the current level of available power or a user-defined energy budget. The key aspects of innovation are methods of selective activation of model components (layers, heads of attention), adaptation of the bit width of calculations and specialized rankaim management for energy consumption management. The requirements for the architecture of the model, the runtime system and the potential benefits in the context of energy efficiency are analyzed. The main technical challenges that require solutions for practical implementation are indicated.