Every Attention Matters: An Efficient Hybrid Architecture for Long-Context Reasoning
This addresses efficiency challenges for AI practitioners working with long-context models, though it appears incremental as it builds on existing attention mechanisms.
The authors tackled the problem of high computational and I/O overhead in long-context reasoning by developing a hybrid architecture integrating linear and softmax attention, resulting in a 10x reduction in inference cost compared to a dense model and over 50% cost reduction compared to prior models, while maintaining SOTA performance on benchmarks.
In this technical report, we present the Ring-linear model series, specifically including Ring-mini-linear-2.0 and Ring-flash-linear-2.0. Ring-mini-linear-2.0 comprises 16B parameters and 957M activations, while Ring-flash-linear-2.0 contains 104B parameters and 6.1B activations. Both models adopt a hybrid architecture that effectively integrates linear attention and softmax attention, significantly reducing I/O and computational overhead in long-context inference scenarios. Compared to a 32 billion parameter dense model, this series reduces inference cost to 1/10, and compared to the original Ring series, the cost is also reduced by over 50%. Furthermore, through systematic exploration of the ratio between different attention mechanisms in the hybrid architecture, we have identified the currently optimal model structure. Additionally, by leveraging our self-developed high-performance FP8 operator library-linghe, overall training efficiency has been improved by 50%. Benefiting from the high alignment between the training and inference engine operators, the models can undergo long-term, stable, and highly efficient optimization during the reinforcement learning phase, consistently maintaining SOTA performance across multiple challenging complex reasoning benchmarks.