It's unofficial implementation of LongRoPE for Llama3.
It's similar as Yarn or RoPE NTK method to support long context LLM However, it is additional beneficial to allow 8 × extension without additional fine-tuning.
Figure 3:LLaMA2-7B perplexity on PG19 and Proof-Pile after extension using different methods, measured without fine-tuning. By fully exploiting the non-uniformities, LongRoPE achieves an 8 × extension without fine-tuning.
that means we could use 64K tokens 8K x 8 = 64K tokens without fine-tuning. So I implement it for evaluate LongRoPE without fine-tune. Authors already mentioned the official site but the official Code is not released yet(24 May2024 ) please check their offical site for official implementation. more detail information, check the paper LongRope :Extending LLM Context Window Beyond 2 Million Tokens
Large context window is a desirable feature in large language models (LLMs). However, due to high fine-tuning costs, scarcity of long texts, and catastrophic values introduced by new token positions, current extended context windows are limited to around 128k tokens.
This paper introduces LongRoPE that, for the first time, extends the context window of pre-trained LLMs to an impressive 2048k tokens, with up to only 1k fine-tuning steps at within 256k training lengths, while maintaining performance at the original short context window. This is achieved by three key innovations: (i) we identify and exploit two forms of non-uniformities in positional interpolation through an efficient search, providing a better initialization for fine-tuning and enabling an 8 × extension in non-fine-tuning scenarios; (ii) we introduce a progressive extension strategy that first fine-tunes a 256k length LLM and then conducts a second positional interpolation on the fine-tuned extended LLM to achieve a 2048k context window; (iii) we readjust LongRoPE on 8k length to recover the short context window performance. Extensive experiments on LLaMA2 and Mistral across various tasks demonstrate the effectiveness of our method. Models extended via LongRoPE retain the original architecture with minor modifications to the positional embedding, and can reuse most pre-existing optimizations. Code will be available at https://github.com/microsoft/LongRoPE
long_rope: longRoPE implementationLongRoPEWrapper: huggingface transformer wrapperexample_llama3_LongRoPE.ipynb: jupyter notebook and generated output samples w/ LongRoPE(8x extention)llama3_eval_baseline.py: simple evaluation code for llama3-8b with baseline RoPEllama3_eval_longRoPE.py: simple evaluation code for llama3-8b with LongRoPE(8x extention)