Activation-Informed Pareto-Guided Low-Rank Compression for Efficient LLM/VLM
The work addresses memory and computing challenges for large models by introducing a low-rank compression framework called PGSVD. It links layer-wise activation errors to overall network loss, proving that uniform error tolerance allocation leads to optimal heterogeneous ranks. This method outperforms existing compression techniques on LLMs and VLMs by achieving significant accuracy improvements while maintaining efficiency.