Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling

MiniCPM Team, Wenhao An, Yingfa Chen, Yewei Fang, Jiayi Li, Xin Li, Yaohui Li, Yishan Li, Yuxuan Li, Biyuan Lin, Chuan Liu, Hezi Liu, Siyuan Liu, Hongya Lyu, Yinxu Pan, Shixin Ren, Xingyu Shen, Zhou Su, Haojun Sun, Yangang Sun, Zhen Leng Thai, Xin Tian, Rui Wang, Xiaorong Wang, Yudong Wang, Bo Wu, Xiaoyue Xu, Dong Xu, Shuaikang Xue, Jiawei Yang, Bowen Zhang, Jinqian Zhang, Letian Zhang, Shengnan Zhang, Xinyu Zhang, Xinyuan Zhang, Zhu Zhang, Hengyu Zhao, Jiacheng Zhao, Jie Zhou, Zihan Zhou, Shuo Wang, Chaojun Xiao, Xu Han, Zhiyuan Liu, Maosong Sun

TL;DR

MiniCPM-SALA is introduced, a 9B-parameter hybrid architecture that integrates the high-fidelity long-context modeling of sparse attention with the global efficiency of linear attention with a hybrid positional encoding (HyPE) to maintain efficiency and performance for long-context tasks.

Abstract

The evolution of large language models (LLMs) towards applications with ultra-long contexts faces challenges posed by the high computational and memory costs of the Transformer architecture. While existing sparse and linear attention mechanisms attempt to mitigate these issues, they typically involve a trade-off between memory efficiency and model performance. This paper introduces MiniCPM-SALA, a 9B-parameter hybrid architecture that integrates the high-fidelity long-context modeling of sparse attention (InfLLM-V2) with the global efficiency of linear attention (Lightning Attention). By employing a layer selection algorithm to integrate these mechanisms in a 1:3 ratio and utilizing a hybrid positional encoding (HyPE), the model maintains efficiency and performance for long-context tasks. Furthermore, we introduce a cost-effective continual training framework that transforms pre-trained Transformer-based models into hybrid models, which reduces training costs by approximately 75% compared to training from scratch. Extensive experiments show that MiniCPM-SALA maintains general capabilities comparable to full-attention models while offering improved efficiency. On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens and supports context lengths of up to 1M tokens, a scale where traditional full-attention 8B models fail because of memory constraints.

MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling

TL;DR

MiniCPM-SALA is introduced, a 9B-parameter hybrid architecture that integrates the high-fidelity long-context modeling of sparse attention with the global efficiency of linear attention with a hybrid positional encoding (HyPE) to maintain efficiency and performance for long-context tasks.

Abstract

The evolution of large language models (LLMs) towards applications with ultra-long contexts faces challenges posed by the high computational and memory costs of the Transformer architecture. While existing sparse and linear attention mechanisms attempt to mitigate these issues, they typically involve a trade-off between memory efficiency and model performance. This paper introduces MiniCPM-SALA, a 9B-parameter hybrid architecture that integrates the high-fidelity long-context modeling of sparse attention (InfLLM-V2) with the global efficiency of linear attention (Lightning Attention). By employing a layer selection algorithm to integrate these mechanisms in a 1:3 ratio and utilizing a hybrid positional encoding (HyPE), the model maintains efficiency and performance for long-context tasks. Furthermore, we introduce a cost-effective continual training framework that transforms pre-trained Transformer-based models into hybrid models, which reduces training costs by approximately 75% compared to training from scratch. Extensive experiments show that MiniCPM-SALA maintains general capabilities comparable to full-attention models while offering improved efficiency. On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens and supports context lengths of up to 1M tokens, a scale where traditional full-attention 8B models fail because of memory constraints.
Paper Structure (10 sections, 3 figures, 4 tables)

This paper contains 10 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Model Development
  3. Model Architecture
  4. Model Training
  5. Experiments
  6. Model Performance
  7. Benchmarks
  8. Inference Speed
  9. Conclusion
  10. Contributions and Acknowledgments

Figures (3)

  • Figure 1: Architecture of MiniCPM-SALA. The model adopts an efficient hybrid design that combines InfLLM-V2 infllmv2 and Lightning Attention lightningattention modules in a 1:3 ratio. Building on an intermediate MiniCPM-4.0 minicpm4 checkpoint, MiniCPM-SALA undergoes a continual training phase to convert a standard Transformer model into a sparse-linear hybrid model.
  • Figure 2: Inference speed comparison between Qwen3-8B and MiniCPM-SALA. For each tested sequence length, the models process a specified input (prefilling) and generate 1K tokens (decoding). "TTFT" denotes Time To First Token, representing the prefilling latency, while "End-to-end" measures the total latency including both prefilling and decoding phases.
  • Figure 3: Inference speed comparison between Qwen3-8B and MiniCPM-SALA. For each tested sequence length, the models process a specified input (prefilling) and generate 1K tokens (decoding).