Submitted by zunhaisu 17 Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus Startlux 6 3