Amazon Science
· Cloud & Big Tech
Making LLMs faster without sacrificing accuracy
A new scaling law that relates particular architectural choices to loss helps identify models that improve throughput by up to 47% with no loss of accuracy.