Activation Sparsity

TEAL: Training-Free Activation Sparsity in Large Language Models

A training-free activation sparsity method for LLM inference that zeros out low-magnitude activations via a quantile threshold, uses a block-wise greedy allocation to protect …