Submitted by zunhaisu 16 Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus Startlux 5 2