H2O evicts non-heavy-hitter tokens from the KV cache using a dynamic submodular policy, retaining recent and frequent-co-occurrence tokens to reduce memory while preserving accuracy.
Improving neural network quantization without retraining using outlier channel splitting
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2023 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
H$_2$O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models
H2O evicts non-heavy-hitter tokens from the KV cache using a dynamic submodular policy, retaining recent and frequent-co-occurrence tokens to reduce memory while preserving accuracy.