Search papers, labs, and topics across Lattice.
This paper introduces KVarN, a novel KV-cache quantization method that addresses the issue of error accumulation in autoregressive decoding of large language models. By employing a calibration-free approach that combines Hadamard rotation with dual-scaling variance normalization, KVarN effectively mitigates the impact of incorrect token scales that typically lead to performance degradation. The results demonstrate that KVarN achieves state-of-the-art performance on generative benchmarks like MATH500, AIME24, and HumanEval at just 2-bit precision, significantly outperforming existing methods.
KVarN reduces error accumulation in autoregressive decoding, setting a new standard for KV-cache quantization with remarkable efficiency at 2-bit precision.
Test-time scaling is a powerful approach to obtain better reasoning in large language models, but it becomes memory-bottlenecked during long-horizon decoding, as the KV-cache grows. KV-cache quantization can help improve this, but current methods are evaluated under prefill-like settings and errors behave differently under autoregressive decoding. We show that in the latter regime, quantization errors accumulate across timesteps, driven primarily by incorrect token scales. We introduce KVarN, a calibration-free KV-cache quantizer that applies a Hadamard rotation followed by a dual-scaling variance normalization across both axes of the K and V matrices. We find that this combination fixes outlying token-scale errors and substantially reduces error accumulation over existing baselines. KVarN establishes a new state-of-theart for KV-cache quantization on generative benchmarks, including MATH500, AIME24 and HumanEval, at 2-bit precision. A vLLM implementation of the KVarN method is available at https://github.com/huawei-csl/KVarN
