Post-transformer inference: 224× compression of Llama-70B with improved accuracy

72 points • anima-core • 3 days ago • 55 comments