Sapling: $\underline{S}$uccessive $\underline{A}$daptation and Com$\underline{p}$ression with $\underline{L}$ayer Dropp$\underline{ing}$ for LLMs
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Efficient Deep Learning, Layer Dropping, LLM Fine-tuning, Specialized LLMs
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We introduce a novel paradigm to fine-tune and compress specialized LLMs with more than 2$\times$ memory saving and inference speedup at deployment time.
Abstract: Specializing Large language models (LLMs) for local deployment and domain-specific use can deliver state-of-the-art performance while meeting latency and privacy requirements. However, conventional task-specific adaptation does not show both memory saving and inference speedup at deployment time. Practical compression techniques like quantization and pruning require hardware support or system optimization to achieve measured inference speedup. We propose Sapling, which can retain LLMs' capacity in a specific knowledge domain and achieve inference speedup on any hardware and deep learning systems by reducing the model depth. Sapling is based on the knowledge localization phenomenon we empirically observed and verified on LLMs, and achieves model compression via successive layer dropping. We evaluated Sapling on LLaMA-7B. At inference time, the models adapted on medical, legal, and financial datasets have all demonstrated reliable performance, comparable memory saving, $1.2$ to $8.5\times$ inference speedup on consumer-level hardware compared to state-of-the-art quantization algorithms, depending on how well the algorithms are supported by efficient accelerator kernels.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2085
Loading