AnTKV: Anchor Token-Aware Sub-Bit Vector Quantization for KV Cache in Large Language Models

Quantization has emerged as an effective and lightweight solution to reduce the memory footprint of the KV cache in Large Language Models (LLMs). Nevertheless, minimizing the performance degradation caused by ultra-low-bit KV cache quantization remains a significant challenge. We observe that quantizing the KV cache of different tokens has varying impacts on the quality of attention outputs. To systematically investigate this phenomenon, we perform forward error propagation analysis on attention and propose the Anchor Score (AnS) that quantifies the sensitivity of each token's KV cache to quantization-induced error. Our analysis reveals significant disparities in AnS across tokens, suggesting that preserving a small subset with full precision (FP16) of high-AnS tokens can greatly mitigate accuracy loss in aggressive quantization scenarios. Based on this insight, we introduce AnTKV, a novel framework that leverages Anchor Token-aware Vector Quantization to compress the KV cache. Furthermore, to support efficient deployment, we design and develop a triton kernel that is fully compatible with FlashAttention, enabling fast online Anchor Token selection. AnTKV enables LLaMA-3-8B to handle context lengths up to 840K tokens on a single 80GB A100 GPU, while achieving up to 3.5x higher decoding throughput compared to the FP16 baseline. Our experiment results demonstrate that AnTKV matches or outperforms prior works such as KIVI, SKVQ, KVQuant, and CQ under 4-bit settings. More importantly, AnTKV achieves significantly lower perplexity under ultra-low-bit quantization on Mistral-7B, with only 6.32 at 1-bit and 8.87 at 0.375-bit, compared to the FP16 baseline of 4.73.

RetroInfer: A Vector-Storage Approach for Scalable Long-Context LLM Inference

The growing context lengths of large language models (LLMs) pose significant challenges for efficient inference, primarily due to GPU memory and bandwidth constraints. We present RetroInfer, a novel system that reconceptualizes the key-value (KV) cache as a vector storage system which exploits the inherent attention sparsity to accelerate long-context LLM inference. At its core is the wave index, an Attention-aWare VEctor index that enables efficient and accurate retrieval of critical tokens through techniques such as tripartite attention approximation, accuracy-bounded attention estimation, and segmented clustering. Complementing this is the wave buffer, which coordinates KV cache placement and overlaps computation and data transfer across GPU and CPU to sustain high throughput. Unlike prior sparsity-based methods that struggle with token selection and hardware coordination, RetroInfer delivers robust performance without compromising model accuracy. Experiments on long-context benchmarks show up to 4.5X speedup over full attention within GPU memory limits and up to 10.5X over sparse attention baselines when KV cache is extended to CPU memory, all while preserving full-attention-level accuracy.

RetrievalAttention: Accelerating Long-Context LLM Inference via Vector Retrieval

Transformer-based large Language Models (LLMs) become increasingly important in various domains. However, the quadratic time complexity of attention operation poses a significant challenge for scaling to longer contexts due to the extremely high inference latency and GPU memory consumption for caching key-value (KV) vectors. This paper proposes RetrievalAttention, a training-free approach to accelerate attention computation. To leverage the dynamic sparse property of attention, RetrievalAttention builds approximate nearest neighbor search (ANNS) indexes upon KV vectors in CPU memory and retrieves the most relevant ones via vector search during generation. Due to the out-of-distribution (OOD) between query vectors and key vectors, off-the-shelf ANNS indexes still need to scan O(N) (usually 30% of all keys) data for accurate retrieval, which fails to exploit the high sparsity. RetrievalAttention first identifies the OOD challenge of ANNS-based attention, and addresses it via an attention-aware vector search algorithm that can adapt to queries and only access 1--3% of data, thus achieving a sub-linear time complexity. RetrievalAttention greatly reduces the inference cost of long-context LLM with much lower GPU memory requirements while maintaining the model accuracy. Especially, RetrievalAttention only needs 16GB GPU memory for serving 128K tokens in LLMs with 8B parameters, which is capable of generating one token in 0.188 seconds on a single NVIDIA RTX4090 (24GB).