Training a SOTA 7B LRM with 1.5x Throughput Improvmentgithub.com/inclusionAI2 pointsjinqueenya year ago