SourceTraining-Free Activation Sparsity in Large Language Models
Using block-wise greedy optimization at 50% model-level sparsity, the paper reveals depth-dependent structure in Llama-3-70B. Attention parameters Wq and Wk exhibit high sparsifiability in block 0 followed by a sharp decline. Wo's sparsifiability starts at 50-60%, peaks at 80-90% in mid-model blocks, then returns to 50-60% in the final blocks. For MLP parameters, Wdown is more sparsifiable than Wgate, which is more sparsifiable than Wup, across all blocks. The authors attribute Wdown's higher sparsifiability to its laplacian-shaped input distribution and Wgate's intermediate position to SiLU decreasing the saliency of negative outputs.