[ad_1]
Authors:
(1) Keivan Alizadeh;
(2) Iman Mirzadeh, Major Contribution;
(3) Dmitry Belenko, Major Contribution;
(4) S. Karen Khatamifard;
(5) Minsik Cho;
(6) Carlo C Del Mundo;
(7) Mohammad Rastegari;
(8) Mehrdad Farajtabar.
Table of Links
2. Flash Memory & LLM Inference and 2.1 Bandwidth and Energy Constraints
3.2 Improving Transfer Throughput with Increased Chunk Sizes
3.3 Optimized Data Management in DRAM
4.1 Results for OPT 6.7B Model
4.2 Results for Falcon 7B Model
6 Conclusion and Discussion, Acknowledgements and References
6 Conclusion and Discussion
In this study, we have tackled the significant challenge of running large language models (LLMs) on devices with constrained memory capacities. Our approach, deeply rooted in the understanding of flash memory and DRAM characteristics, represents a novel convergence of hardware-aware strategies and machine learning. By developing an inference cost model that aligns with these hardware constraints, we have introduced two innovative techniques: ’windowing’ and ’row-column bundling.’ These methods collectively contribute to a significant reduction in the data load and an increase in the efficiency of memory usage. Weight bundling and windowing are two very basic techniques aimed at showcasing the potentials to increase chunk size and read sequentiality while reducing data transfer through sparsity. Numerous opportunities exist for developing smarter and more efficient methods to achieve these objectives.
The practical outcomes of our research are noteworthy. We have demonstrated the ability to run LLMs up to twice the size of available DRAM, achieving an acceleration in inference speed by 4-5x compared to traditional loading methods in CPU, and 20-25x in GPU. This innovation is particularly crucial for deploying advanced LLMs in resource-limited environments, thereby expanding their applicability and accessibility. The PyTorch based implementation for forward pass have only undergone algorithmic (as opposed to systems) optimization. Significant additional gains are expected from a custom lower level implementation.
Our work not only provides a solution to a current computational bottleneck but also sets a precedent for future research. It underscores the importance of considering hardware characteristics in the development of inference-optimized algorithms, suggesting a promising direction for further explorations in this domain. We believe as LLMs continue to grow in size and complexity, approaches like this work will be essential for harnessing their full potential in a wide range of devices and applications.
Our study represents an initial endeavor in the pursuit of democratizing Large Language Model (LLM) inference, making it accessible to a wider array of individuals and devices. We recognize that this early effort has its limitations, which, in turn, open up compelling avenues for future research. A critical aspect for future exploration is the analysis of power consumption and thermal limitations inherent in the methods we propose, particularly for on-device deployment. Currently, our focus is on single-batch inference. However, expanding this to include scenarios like prompt processing, multi-batch inference, and speculative decoding presents itself as a valuable area for further investigation. In our initial proof of concept, we operated under the assumption of memory availability being half the size of the model. Exploring the dynamics of working with varying memory sizes—both larger and smaller—introduces a fascinating balance between latency and accuracy, and is a compelling area for future exploration. In conclusion, our methodology is constructed on the foundation of sparsified networks. Nonetheless, the underlying concept holds potential for broader applications. It can be adapted to selectively load weights in nonsparse networks or to dynamically retrieve model weights from flash storage. This adaptation would be contingent on the specific requirements of the input prompt or the contextual parameters provided. Such an approach suggests a versatile strategy for managing model weights, optimizing performance based on the nature of the input, thereby enhancing the efficiency, usefulness, and applicability of the proposed scheme in various scenarios dealing with Large Language Models (LLMs).
Acknowledgements
We would like to thank Itay Sagron, Lailin Chen, Mahyar Najibi, Qichen Fu, Moin Nabi, Peter Zatloukal, Arsalan Farooq, Sachin Mehta, Mohammad Samragh, Matt Johnson, Etai Zaltsman, Lin Chang, Dominic Giampaolo, Taal Uliel, Hadi Pouransari, Fartash Faghri, Oncel Tuzel, Samy Bengio, Ruoming Pang, Chong Wang, Ronan Collobert, David Grangier, and Aftab Munshi for the valuable feedback and discussions.
References
Udit Agrawal, Rangharajan Venkatesan, Brucek Khailany, Stephen W Keckler, and William J Dally. 2022. Atomlayer: minimizing dram data movement for ultra-sparse models on gpus. In Proceedings of the 27th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, pages 223–238.
Arash Ahmadian, Saurabh Dash, Hongyu Chen, Bharat Venkitesh, Stephen Gou, Phil Blunsom, A. Ustun, and Sara Hooker. 2023. Intriguing properties of quantization at scale. ArXiv, abs/2305.19268.
Ebtesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Maitha Alhammadi, Mazzotta Daniele, Daniel Heslow, Julien Launay, Quentin Malartic, Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. 2023. The falcon series of language models: Towards open frontier models.
Reza Yazdani Aminabadi, Samyam Rajbhandari, Ammar Ahmad Awan, Cheng Li, Du Li, Elton Zheng, Olatunji Ruwase, Shaden Smith, Minjia Zhang, Jeff Rasley, et al. 2022. Deepspeed-inference: enabling efficient inference of transformer models at unprecedented scale. In SC22: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1–15. IEEE.
Sangmin Bae, Jongwoo Ko, Hwanjun Song, and Se-Young Yun. 2023. Fast and robust earlyexiting framework for autoregressive language models with synchronized parallel decoding. ArXiv, abs/2310.05424.
Cenk Baykal, Dylan Cutler, Nishanth Dikkala, Nikhil Ghosh, Rina Panigrahy, and Xin Wang. 2023. Alternating updates for efficient transformers. ArXiv, abs/2301.13310.
Alexander Borzunov, Dmitry Baranchuk, Tim Dettmers, Maksim Riabinin, Younes Belkada, Artem Chumachenko, Pavel Samygin, and Colin Raffel. 2023. Petals: Collaborative inference and fine-tuning of large models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 558–568, Toronto, Canada. Association for Computational Linguistics.
Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901.
Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311.
Han Dai, Yi Zhang, Ziyu Gong, Nanqing Yang, Wei Dai, Eric Song, and Qiankun Xie. 2021. Spatten: Efficient sparse attention architecture with cascade token and head pruning. In Advances in Neural Information Processing Systems, volume 34.
Erich Elsen, Augustus Odena, Maxwell Nye, Sag-˘ nak Ta¸sırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, and Arushi Somani. 2023. Releasing Persimmon-8B.
Mingyu Gao, Jie Yu, Wentai Li, Michael C Dai, Nam Sung Kim, and Krste Asanovic. 2022. computedram: In-memory compute using off-the-shelf dram. In Proceedings of the 27th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, pages 1065– 1079.
Alex Graves. 2016. Adaptive computation time for recurrent neural networks. In International Conference on Machine Learning, pages 3500–3509. PMLR.
Jongmin Ham, Jinha Kim, Jinwoong Choi, Cheolwoo Cho, Seulki Hong, Kyeongsu Han, and Taejoo Chung. 2016. Graphssd: a high performance flash-based storage system for large-scale graph processing. In 2016 USENIX Annual Technical Conference (USENIXATC 16), pages 243–256.
Song Han, Xingyu Liu, Huizi Mao, Jing Pu, Ardavan Pedram, Mark A Horowitz, and William J Dally. 2016a. Eie: efficient inference engine on compressed deep neural network. arXiv preprint arXiv:1602.01528.
Song Han, Huizi Mao, and William J Dally. 2016b. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In International Conference on Learning Representations (ICLR).
Zhenyu He, Zexuan Zhong, Tianle Cai, Jason D Lee, and Di He. 2023. Rest: Retrieval-based speculative decoding. ArXiv, abs/2311.08252.
Duc Nien Hoang, Minsik Cho, Thomas Merth, Mohammad Rastegari, and Zhangyang Wang. 2023. (dynamic) prompting might be all you need to repair compressed llms. ArXiv, abs/2310.00867.
Ajay Jaiswal, Zhe Gan, Xianzhi Du, Bowen Zhang, Zhangyang Wang, and Yinfei Yang. 2023. Compressing llms: The truth is rarely pure and never simple. ArXiv, abs/2310.01382.
Yaniv Leviathan, Matan Kalman, and Yossi Matias. 2022. Fast inference from transformers via speculative decoding.
Jiaxi Li and Wei Lu. 2023. Contextual distortion reveals constituency: Masked language models are implicit parsers. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5208–5222, Toronto, Canada. Association for Computational Linguistics.
Liang Li, Qingyuan Li, Bo Zhang, and Xiangxiang Chu. 2023. Norm tweaking: High-performance lowbit quantization of large language models. ArXiv, abs/2309.02784.
Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Xingyu Dang, and Song Han. 2023. Awq: Activationaware weight quantization for llm compression and acceleration. ArXiv, abs/2306.00978.
Zechun Liu, Barlas Oguz, Changsheng Zhao, Ernie ˘ Chang, Pierre Stock, Yashar Mehdad, Yangyang Shi, Raghuraman Krishnamoorthi, and Vikas Chandra. 2023a. Llm-qat: Data-free quantization aware training for large language models. ArXiv, abs/2305.17888.
Zichang Liu, Jue Wang, Tri Dao, Tianyi Zhou, Binhang Yuan, Zhao Song, Anshumali Shrivastava, Ce Zhang, Yuandong Tian, Christopher Re, et al. 2023b. Deja vu: Contextual sparsity for efficient llms at inference time. In International Conference on Machine Learning, pages 22137–22176. PMLR.
Moinuddin K Meswani, Sergey Blagodurov, David Roberts, John Slice, Mike Ignatowski, and Gabriel Loh. 2015. Neural cache: Bit-serial in-cache acceleration of deep neural networks. In 2015 48th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 383–394. IEEE.
Iman Mirzadeh, Keivan Alizadeh, Sachin Mehta, Carlo C Del Mundo, Oncel Tuzel, Golnoosh Samei, Mohammad Rastegari, and Mehrdad Farajtabar. 2023. Relu strikes back: Exploiting activation sparsity in large language models.
Sharan Narang, Logan Feistel, Erich Elsen Undersander, Cindy Song, and Gregory Diamos. 2022. Firefly: A lightweight system for running multi-billion parameter models on commodity hardware. In 2022 ACM/IEEE 49th Annual International Symposium on Computer Architecture (ISCA), pages 757–771. IEEE.
Sharan Narang, Erich Elsen Undersander, and Gregory Diamos. 2021. Sparse gpu kernels for deep learning. In International Conference on Learning Representations.
Angshuman Parashar, Minsoo Rhu, Anurag Mukkara, Antonio Puglielli, Rangharajan Venkatesan, Brucek Khailany, Joel Emer, Stephen W Keckler, and William J Dally. 2017. Timeloop: A systematic approach to dnn accelerator evaluation. In 2017 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), pages 241–251. IEEE.
Samyam Rajbhandari, Olatunji Ruwase, Jeff Rasley, Shaden Smith, and Yuxiong He. 2021. Zero-infinity: Breaking the gpu memory wall for extreme scale deep learning. In SC21: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1–14.
Minsoo Rhu, Natalia Gimelshein, Jason Clemons, Arslan Zulfiqar, and Stephen W Keckler. 2013. vdnn: Virtualized deep neural networks for scalable, memory-efficient neural network design. In 2016 49th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), page Article 13. IEEE Computer Society.
Wenqi Shao, Mengzhao Chen, Zhaoyang Zhang, Peng Xu, Lirui Zhao, Zhiqiang Li, Kaipeng Zhang, Peng Gao, Yu Jiao Qiao, and Ping Luo. 2023. Omniquant: Omnidirectionally calibrated quantization for large language models. ArXiv, abs/2308.13137.
Yifan Shao, Mengjiao Li, Wenhao Cai, Qi Wang, Dhananjay Narayanan, and Parthasarathy Ranganathan. 2022. Hotpot: Warmed-up gigascale inference with tightly-coupled compute and reuse in flash. In Proceedings of the 55th Annual IEEE/ACM International Symposium on Microarchitecture, pages 335–349.
Ying Sheng, Lianmin Zheng, Binhang Yuan, Zhuohan Li, Max Ryabinin, Beidi Chen, Percy Liang, Christopher Ré, Ion Stoica, and Ce Zhang. 2023. Flexgen: High-throughput generative inference of large language models with a single GPU. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research, pages 31094–31116. PMLR.
Vedant Subramani, Marios Savvides, Li Ping, and Sharan Narang. 2022. Adapt: Parameter adaptive tokenwise inference for vision transformers. In Proceedings of the 55th Annual IEEE/ACM International Symposium on Microarchitecture.
Mingjie Sun, Zhuang Liu, Anna Bair, and J. Zico Kolter. 2023. A simple and effective pruning approach for large language models. ArXiv, abs/2306.11695.
Haojun Xia, Zhen Zheng, Yuchao Li, Donglin Zhuang, Zhongzhu Zhou, Xiafei Qiu, Yong Li, Wei Lin, and Shuaiwen Leon Song. 2023. Flash-llm: Enabling low-cost and highly-efficient large generative model inference with unstructured sparsity. Proc. VLDB Endow., 17:211–224.
Zhaozhuo Xu, Zirui Liu, Beidi Chen, Yuxin Tang, Jue Wang, Kaixiong Zhou, Xia Hu, and Anshumali Shrivastava. 2023. Compress, then prompt: Improving accuracy-efficiency trade-off of llm inference with transferable prompt. ArXiv, abs/2305.11186.
Rongjie Yi, Liwei Guo, Shiyun Wei, Ao Zhou, Shangguang Wang, and Mengwei Xu. 2023. Edgemoe: Fast on-device inference of moe-based large language models. ArXiv, abs/2308.14352.
Jinchao Zhang, Jue Wang, Huan Li, Lidan Shou, Ke Chen, Gang Chen, and Sharad Mehrotra. 2023. Draft & verify: Lossless large language model acceleration via self-speculative decoding. ArXiv, abs/2309.08168.
Shizhao Zhang, Han Dai, Tian Sheng, Jiawei Zhang, Xiaoyong Li, Qun Xu, Mengjia Dai, Yunsong Xiao, Chao Ma, Rui Tang, et al. 2022a. Llm quantization: Quantization-aware training for large language models. In Advances in Neural Information Processing Systems, volume 35.
Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona T. Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. 2022b. OPT: open pre-trained transformer language models. CoRR, abs/2205.01068.
Yilong Zhao, Chien-Yu Lin, Kan Zhu, Zihao Ye, Lequn Chen, Size Zheng, Luis Ceze, Arvind Krishnamurthy, Tianqi Chen, and Baris Kasikci. 2023. Atom: Lowbit quantization for efficient and accurate llm serving. ArXiv, abs/2310.19102.
[ad_2]
Source link