1 paper · 1 filter
Liza Babaoglu, Shuangyi Chen, Ashish Khisti
As large language models (LLMs) are increasingly deployed across heterogeneous hardware with varying resource constraints, the ability to adaptively manage the performance-efficien…