LaMPE: Length-aware Multi-grained Positional Encoding for Adaptive Long-context Scaling Without Training

Open in new window