LongRoPE2: Near-Lossless LLM Context Window Scaling

TL;DR

LongRoPE2 introduces a novel method to extend large language models' context windows to 128K tokens while maintaining high performance on shorter contexts, using efficient training and rescaling techniques.

Contribution

It presents a new RoPE rescaling algorithm and a mixed context training approach that significantly extend context length with minimal additional training data.

Findings

Achieves 128K effective context length on LLaMA3-8B.

Retains over 98.5% of short-context performance.

Requires only 10B tokens for training, much less than previous methods.

Abstract

LongRoPE2 is a novel approach that extends the effective context window of pre-trained large language models (LLMs) to the target length, while preserving the performance on the original shorter context window. This is achieved by three contributions: (1) a hypothesis that insufficient training in higher RoPE dimensions contributes to the persistent out-of-distribution (OOD) issues observed in existing methods; (2) an effective RoPE rescaling algorithm that adopts evolutionary search guided by "needle-driven" perplexity to address the insufficient training problem; (3) a mixed context window training approach that fine-tunes model weights to adopt rescaled RoPE for long-context sequences while preserving the short-context performance with the original RoPE. Extensive experiments on LLaMA3-8B and Phi3-mini-3.8B across various benchmarks validate the hypothesis and demonstrate the effectiveness of LongRoPE2. Remarkably, LongRoPE2 extends LLaMA3-8B to achieve a 128K effective context length while retaining over 98.5% of short-context performance, using only 10B tokens -- 80x fewer than Meta's approach, which fails to reach the target effective context length. Code will be available at https://github.com/microsoft/LongRoPE.