RaBitQCache: Rotated Binary Quantization for KVCache in Long Context LLM Inference
About
Long-context Large Language Model inference is severely bottlenecked by the massive Key-Value (KV) cache, yet existing sparse attention methods often suffer from static fixed-budget (Top-k) retrieval or rely on proxy scores that are computationally expensive and biased. To address these limitations, we propose RaBitQCache, a novel sparse attention framework that utilizes randomized rotated binary quantization and high-throughput binary-INT4 arithmetic to efficiently estimate attention weights. Our proxy score serves as an unbiased estimator with a proven error bound, enabling adaptive Top-p retrieval that dynamically adjusts the token budget based on actual attention sparsity. We further implement a hardware-aware system with asynchronous pipelining and lazy updates to mask overhead. Evaluations demonstrate that RaBitQCache significantly accelerates inference and reduces memory I/O while preserving generation quality compared to state-of-the-art baselines. Code is available at https://github.com/Sakuraaa0/RaBitQCache.git.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Long-context Language Understanding | LongBench | Average Score54.58 | 90 | |
| Long-context language modeling | RULER | Accuracy (8K Context)84.7 | 80 | |
| Long-context language modeling | LongBench | Generation Score50.6 | 24 | |
| Mathematical Reasoning | GSM8K | Accuracy77 | 13 | |
| Long-context Language Understanding | LongBench | Average Score50.63 | 6 | |
| Mathematical Reasoning | GSM8K | Accuracy (GSM8K)77 | 4 | |
| Synthetic Long-context Retrieval | RULER 8K–64K | Average Accuracy79.55 | 4 |