Papers with FlashAttention
QiMeng-Attention: SOTA Attention Operator is generated by SOTA Attention Algorithm (2025.findings-acl)
Copied to clipboard
Qirui Zhou, Shaohui Peng, Weiqiang Xiong, Haixin Chen, Yuanbo Wen, Haochen Li, Ling Li, Qi Guo, Yongwei Zhao, Ke Gao, Ruizhi Chen, Yanjun Wu, Zhao Chen, Yunji Chen
| Challenge: | Existing LLMs cannot comprehend the complex data flow and computation process of the attention operator and utilize low-level primitive to exploit GPU performance. |
| Approach: | They propose an LLM-friendly Thinking Language (LLM-TL) that can decouple the generation of high-level optimization logic and low-level implementation on GPU and enhance LLMs’ understanding of attention operator. |
| Outcome: | The proposed method outshines existing LLMs on A100, RTX8000, and T4 GPUs, achieving a speed-up of up to 35.16. |
OjaKV: Context-Aware Online Low-Rank KV Cache Compression (2026.findings-acl)
Copied to clipboard
Yuxuan Zhu, David H. Yang, Mohammad Mohammadi Amiri, Keerthiram Murugesan, Tejaswini Pedapati, Pin-Yu Chen
| Challenge: | Existing methods for inference use static, offline-learned subspaces that perform poorly under distribution shifts. |
| Approach: | They propose a framework that integrates a storage policy with an online subspace adaptation to preserve key-value tokens in full rank as high-fidelity anchors. |
| Outcome: | Experiments show that OjaKV maintains or improves zero-shot accuracy at high compression ratios, achieving the strongest gains on long-context benchmarks requiring complex reasoning. |
MiniKV: Pushing the Limits of 2-Bit KV Cache via Compression and System Co-Design for Efficient Long Context Inference (2025.findings-acl)
Copied to clipboard
| Challenge: | State-of-the-art 2-bit KV cache quantization methods achieve excellent results in accelerating LLM inference while retaining accuracy on long context tasks. |
| Approach: | They propose a method based on 2-bit KV cache quantization with adaptive KV policies that retain LLM accuracy with only a subset of KV states. |
| Outcome: | The proposed method outperforms state-of-the-art methods on a wide range of long context tasks while retaining accuracy. |
A Simple and Effective L_2 Norm-Based Strategy for KV Cache Compression (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to reduce the KV cache size involve fine-tuning the model to learn a compression strategy or leveraging attention scores to reduce sequence length. |
| Approach: | They find a correlation between the L2 norm and attention scores over cached KV pairs . they compress the KV cache based on the L1 norm of key embeddings . |
| Outcome: | The proposed approach reduces the KV cache size by 50% on language modelling and needle-in-a-haystack tasks and 90% on passkey retrieval tasks without losing accuracy. |
Stability Implies Redundancy: Delta Attention Selective Halting for Efficient Long-Context Prefilling (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods to reduce sequence length rely on heuristics that break compatibility with hardware-efficient kernels like FlashAttention. |
| Approach: | They propose a method that selectively halts stabilized tokens by monitoring layer-wise update dynamics of the self-attention mechanism. |
| Outcome: | The proposed method can reduce prefill complexity while preserving model accuracy and hardware efficiency. |
SpecExtend: A Drop-in Enhancement for Speculative Decoding of Long Sequences (2026.findings-acl)
Copied to clipboard
| Challenge: | Speculative decoding performance degrades as input length increases, with significant drops even at moderate lengths. |
| Approach: | They propose a drop-in enhancement that improves speculative decoding on long sequences without additional training. |
| Outcome: | The proposed enhancement accelerates speculative decoding by up to 2.84 on 16K-token long document summarization and up to 3.86 on long-form reasoning while preserving the short-input performance of state-of-the-art frameworks. |