Researchers baked 3x inference speedups directly into LLM weights — without speculative decoding

Researchers from the University of Maryland, Lawrence Livermore, Columbia and TogetherAI have developed a training technique that triples LLM inference speed without auxiliary models or infrastructure ...
IEEE

CQS-Attention: Scaling Up the Standard Attention Computation for Infinitely Long Sequences

Abstract: Transformer models suffer from unaffordable high memory consumption when the sequence is long and standard self-attention is utilized. We developed a sequence parallelism scheme called ...
IEEE

A Sequence Fine-Tuning Strategy Based Evolutionary Algorithm for Solving Project Scheduling Problems

Abstract: As artificial intelligence technology advances, the industrial landscape has been gradually transitioning from man-ual labor to automated manufacturing processes. This shift has highlighted ...