Research map / Large language models / Retrieval and knowledge
LazyAttention: Efficient Retrieval-Augmented Generation with Deferred Positional Encoding
Haocheng Xia, Mihir Pamnani, Hanxi Fang, Supawit Chockchowwat, Yongjoo Park
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| defer RoPE into attention kernel for position-agnostic KV reuse | Block-Attention (vLLM) · skewed document traffic | time-to-first-token: 1.37x lower TTFT | Improved |
| defer RoPE into attention kernel for position-agnostic KV reuse | Block-Attention (vLLM) · skewed document traffic | inference throughput: 1.40x higher | Improved |
| store single position-agnostic KV copy per document | Block-Attention (vLLM) · 1 GB cache, high skew | VRAM cache hit ratio: 7.27 -> 13.57 | Improved |
| store single position-agnostic KV copy per document | Block-Attention (vLLM) · no-limit cache, mid skew | VRAM cache hit ratio: 27.38 -> 29.09 | Improved |
| single relative rotation of K inside fused kernel instead of two rotations (ablation) | naive deferred encoding · prefilling kernel | FLOPs overhead: 3/(2M) -> 3/(4M); 0.59% at M=128 | Improved |
| rotate Q only when document offset changes in decoding (ablation) | LazyAttention decoding kernel · decoding | per-token overhead: 0.13% overhead | Improved |
| use LazyAttention instead of Block-Attention | Tulu3-Block-FT · QA average exact match | accuracy: 71.2 -> 69.2 | No clear effect |
| use LazyAttention instead of CacheBlend | Tulu3-Block-FT · QA average exact match | accuracy: 67.5 -> 69.2 | Improved |
| use deferred rotation under uniform traffic | LazyAttention · Uniform document sampling | TTFT | No clear effect |
Closest papers by meaning
- RazorAttention: Efficient KV Cache Compression Through Retrieval Heads (ICLR 2025 · similarity 0.91)
- KVLink: Accelerating Large Language Models via Efficient KV Cache Reuse (NeurIPS 2025 · similarity 0.89)
- EPIC: Efficient Position-Independent Caching for Serving Large Language Models (ICML 2025 · similarity 0.88)
- RetrievalAttention: Accelerating Long-Context LLM Inference via Vector Retrieval (NeurIPS 2025 · similarity 0.88)
- DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads (ICLR 2025 · similarity 0.88)
- ArkVale: Efficient Generative LLM Inference with Recallable Key-Value Eviction (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
