ASVD compresses LLMs by 10-30% and KV caches by 50% via activation-aware SVD that absorbs outliers into transformed weights and calibrates per-layer sensitivity.
Low- rank matrix factorization for deep neural network training with high-dimensional output targets
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2023 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
ASVD: Activation-aware Singular Value Decomposition for Compressing Large Language Models
ASVD compresses LLMs by 10-30% and KV caches by 50% via activation-aware SVD that absorbs outliers into transformed weights and calibrates per-layer sensitivity.