Farewell to Length Extrapolation, a Training-Free Infinite Context with Finite Attention Scope

Open in new window