Hierarchical Context Merging: Better Long Context Understanding for Pre-trained LLMs
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2024
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866914756688871424 |
|---|---|
| author | Song, Woomin Oh, Seunghyuk Mo, Sangwoo Kim, Jaehyung Yun, Sukmin Ha, Jung-Woo Shin, Jinwoo |
| author_facet | Song, Woomin Oh, Seunghyuk Mo, Sangwoo Kim, Jaehyung Yun, Sukmin Ha, Jung-Woo Shin, Jinwoo |
| contents | Large language models (LLMs) have shown remarkable performance in various natural language processing tasks. However, a primary constraint they face is the context limit, i.e., the maximum number of tokens they can process. Previous works have explored architectural changes and modifications in positional encoding to relax the constraint, but they often require expensive training or do not address the computational demands of self-attention. In this paper, we present Hierarchical cOntext MERging (HOMER), a new training-free scheme designed to overcome the limitations. HOMER uses a divide-and-conquer algorithm, dividing long inputs into manageable chunks. Each chunk is then processed collectively, employing a hierarchical strategy that merges adjacent chunks at progressive transformer layers. A token reduction technique precedes each merging, ensuring memory usage efficiency. We also propose an optimized computational order reducing the memory requirement to logarithmically scale with respect to input length, making it especially favorable for environments with tight memory restrictions. Our experiments demonstrate the proposed method's superior performance and memory efficiency, enabling the broader use of LLMs in contexts requiring extended context. Code is available at https://github.com/alinlab/HOMER. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2404_10308 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Hierarchical Context Merging: Better Long Context Understanding for Pre-trained LLMs Song, Woomin Oh, Seunghyuk Mo, Sangwoo Kim, Jaehyung Yun, Sukmin Ha, Jung-Woo Shin, Jinwoo Machine Learning Artificial Intelligence Large language models (LLMs) have shown remarkable performance in various natural language processing tasks. However, a primary constraint they face is the context limit, i.e., the maximum number of tokens they can process. Previous works have explored architectural changes and modifications in positional encoding to relax the constraint, but they often require expensive training or do not address the computational demands of self-attention. In this paper, we present Hierarchical cOntext MERging (HOMER), a new training-free scheme designed to overcome the limitations. HOMER uses a divide-and-conquer algorithm, dividing long inputs into manageable chunks. Each chunk is then processed collectively, employing a hierarchical strategy that merges adjacent chunks at progressive transformer layers. A token reduction technique precedes each merging, ensuring memory usage efficiency. We also propose an optimized computational order reducing the memory requirement to logarithmically scale with respect to input length, making it especially favorable for environments with tight memory restrictions. Our experiments demonstrate the proposed method's superior performance and memory efficiency, enabling the broader use of LLMs in contexts requiring extended context. Code is available at https://github.com/alinlab/HOMER. |
| title | Hierarchical Context Merging: Better Long Context Understanding for Pre-trained LLMs |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2404.10308 |