Hybrid Linear Attention Done Right: Efficient Distillation and Effective Architectures for Extremely Long Contexts

TL;DR

This paper introduces HALO, a distillation pipeline, and HypeNet, a hybrid architecture with novel position encoding, enabling efficient long-context modeling with minimal additional training data.

Contribution

The paper proposes a new distillation method and architecture that significantly improve long-context performance and efficiency of hybrid models with minimal data.

Findings

HypeNet achieves performance comparable to original Transformers.

HALO requires only 2.3B tokens for conversion, less than 0.01% of pre-training data.

HypeNet demonstrates superior length generalization and efficiency.

Abstract

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