Hybrid Linear Attention Done Right: Efficient Distillation and Effective Architectures for Extremely Long Contexts
This work addresses the high computational barrier for adopting hybrid models in long-context AI applications, offering a more accessible solution for researchers and practitioners.
The paper tackles the prohibitive cost of pre-training hybrid Transformer-RNN models for long contexts by introducing HALO for efficient distillation and HypeNet with novel position encoding, achieving comparable performance to original Transformers with only 2.3B tokens (less than 0.01% of pre-training data) and superior long-context efficiency.
Hybrid Transformer architectures, which combine softmax attention blocks and recurrent neural networks (RNNs), have shown a desirable performance-throughput tradeoff for long-context modeling, but their adoption and studies are hindered by the prohibitive cost of large-scale pre-training from scratch. Some recent studies have shown that pre-trained softmax attention blocks can be converted into RNN blocks through parameter transfer and knowledge distillation. However, these transfer methods require substantial amounts of training data (more than 10B tokens), and the resulting hybrid models also exhibit poor long-context performance, which is the scenario where hybrid models enjoy significant inference speedups over Transformer-based models. In this paper, we present HALO (Hybrid Attention via Layer Optimization), a pipeline for distilling Transformer models into RNN-attention hybrid models. We then present HypeNet, a hybrid architecture with superior length generalization enabled by a novel position encoding scheme (named HyPE) and various architectural modifications. We convert the Qwen3 series into HypeNet using HALO, achieving performance comparable to the original Transformer models while enjoying superior long-context performance and efficiency. The conversion requires just 2.3B tokens, less than 0.01% of their pre-training data