Submitted by zunhaisu 24 Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus Startlux 7 4