KVTuner: Sensitivity-Aware Layer-Wise Mixed-Precision KV Cache Quantization for Efficient and Nearly Lossless LLM Inference

Xing Li, Zeyu Xing, Yiming Li, Linping Qu, Hui-Ling Zhen, Yiwu Yao, Wulong Liu, Sinno Jialin Pan, Mingxuan Yuan
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:36451-36485, 2025.

Abstract

KV cache quantization can improve Large Language Models (LLMs) inference throughput and latency in long contexts and large batch-size scenarios while preserving LLMs effectiveness. However, current methods have three unsolved issues: overlooking layer-wise sensitivity to KV cache quantization, high overhead of online fine-grained decision-making, and low flexibility to different LLMs and constraints. Therefore, we theoretically analyze the inherent correlation of layer-wise transformer attention patterns to KV cache quantization errors and study why key cache is generally more important than value cache for quantization error reduction. We further propose a simple yet effective framework KVTuner to adaptively search for the optimal hardware-friendly layer-wise KV quantization precision pairs for coarse-grained KV cache with multi-objective optimization and directly utilize the offline searched configurations during online inference. To reduce the computational cost of offline calibration, we utilize the intra-layer KV precision pair pruning and inter-layer clustering to reduce the search space. Experimental results show that we can achieve nearly lossless 3.25-bit mixed precision KV cache quantization for LLMs like Llama-3.1-8B-Instruct and 4.0-bit for sensitive models like Qwen2.5-7B-Instruct on mathematical reasoning tasks. The maximum inference throughput can be improved by 21.25% compared with KIVI-KV8 quantization over various context lengths. Our code and searched configurations are available at https://github.com/cmd2001/KVTuner.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-li25dd, title = {{KVT}uner: Sensitivity-Aware Layer-Wise Mixed-Precision {KV} Cache Quantization for Efficient and Nearly Lossless {LLM} Inference}, author = {Li, Xing and Xing, Zeyu and Li, Yiming and Qu, Linping and Zhen, Hui-Ling and Yao, Yiwu and Liu, Wulong and Pan, Sinno Jialin and Yuan, Mingxuan}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {36451--36485}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/li25dd/li25dd.pdf}, url = {https://proceedings.mlr.press/v267/li25dd.html}, abstract = {KV cache quantization can improve Large Language Models (LLMs) inference throughput and latency in long contexts and large batch-size scenarios while preserving LLMs effectiveness. However, current methods have three unsolved issues: overlooking layer-wise sensitivity to KV cache quantization, high overhead of online fine-grained decision-making, and low flexibility to different LLMs and constraints. Therefore, we theoretically analyze the inherent correlation of layer-wise transformer attention patterns to KV cache quantization errors and study why key cache is generally more important than value cache for quantization error reduction. We further propose a simple yet effective framework KVTuner to adaptively search for the optimal hardware-friendly layer-wise KV quantization precision pairs for coarse-grained KV cache with multi-objective optimization and directly utilize the offline searched configurations during online inference. To reduce the computational cost of offline calibration, we utilize the intra-layer KV precision pair pruning and inter-layer clustering to reduce the search space. Experimental results show that we can achieve nearly lossless 3.25-bit mixed precision KV cache quantization for LLMs like Llama-3.1-8B-Instruct and 4.0-bit for sensitive models like Qwen2.5-7B-Instruct on mathematical reasoning tasks. The maximum inference throughput can be improved by 21.25% compared with KIVI-KV8 quantization over various context lengths. Our code and searched configurations are available at https://github.com/cmd2001/KVTuner.} }
Endnote
%0 Conference Paper %T KVTuner: Sensitivity-Aware Layer-Wise Mixed-Precision KV Cache Quantization for Efficient and Nearly Lossless LLM Inference %A Xing Li %A Zeyu Xing %A Yiming Li %A Linping Qu %A Hui-Ling Zhen %A Yiwu Yao %A Wulong Liu %A Sinno Jialin Pan %A Mingxuan Yuan %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-li25dd %I PMLR %P 36451--36485 %U https://proceedings.mlr.press/v267/li25dd.html %V 267 %X KV cache quantization can improve Large Language Models (LLMs) inference throughput and latency in long contexts and large batch-size scenarios while preserving LLMs effectiveness. However, current methods have three unsolved issues: overlooking layer-wise sensitivity to KV cache quantization, high overhead of online fine-grained decision-making, and low flexibility to different LLMs and constraints. Therefore, we theoretically analyze the inherent correlation of layer-wise transformer attention patterns to KV cache quantization errors and study why key cache is generally more important than value cache for quantization error reduction. We further propose a simple yet effective framework KVTuner to adaptively search for the optimal hardware-friendly layer-wise KV quantization precision pairs for coarse-grained KV cache with multi-objective optimization and directly utilize the offline searched configurations during online inference. To reduce the computational cost of offline calibration, we utilize the intra-layer KV precision pair pruning and inter-layer clustering to reduce the search space. Experimental results show that we can achieve nearly lossless 3.25-bit mixed precision KV cache quantization for LLMs like Llama-3.1-8B-Instruct and 4.0-bit for sensitive models like Qwen2.5-7B-Instruct on mathematical reasoning tasks. The maximum inference throughput can be improved by 21.25% compared with KIVI-KV8 quantization over various context lengths. Our code and searched configurations are available at https://github.com/cmd2001/KVTuner.
APA
Li, X., Xing, Z., Li, Y., Qu, L., Zhen, H., Yao, Y., Liu, W., Pan, S.J. & Yuan, M.. (2025). KVTuner: Sensitivity-Aware Layer-Wise Mixed-Precision KV Cache Quantization for Efficient and Nearly Lossless LLM Inference. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:36451-36485 Available from https://proceedings.mlr.press/v267/li25dd.html.

Related Material