Megalodon also uses “chunk-wise attention,” which divides the input sequence into fixed-size blocks to reduce the complexity of the model from quadratic to linear.Read More
Ben Dickson
Source link
Meta challenges transformer architecture with Megalodon LLM
