Nicolò Felicioni

New method uses idle inference resources to cut LLM training costs

Researchers have developed a novel method to reduce the computational cost of training large language models by utilizing idle inference resources. The algorithm predicts gradients using a reduced-precision, reverse-mod…