This AI Paper Demonstrates An End-to-End Training Flow on An Large Language Model LLM-13 Billion GPT-Using Sparsity And Dataflow

Machine learning system implementation in the academic and commercial domains has been expedited by foundation models in the natural language processing and computer vision domains. Researchers have suggested increasing parameter count by orders of magnitude to extract additional capabilities from these models and train on vast data corpora. Their primary traits of self-regulation and adaptability enable a wide range of applications to be developed to address particular issues, including text production, sentiment analysis, picture segmentation, and image recognition. 

Due to power and physical limitations, the underlying hardware used to train such enormous models needs to scale proportionally to model parameters. Several techniques have been investigated to overcome this computational challenge, including network restructuring, network pruning, network quantization, low-rank decomposition knowledge distillation, model sparsity, etc. Different types of sparse approaches have been put forth to lower computing intensity and imitate the connections between neurons in the human brain. The underlying hardware architecture presents new difficulties as sparsity methods advance and becomes widely used in training and inference applications. 

A well-balanced system needs to tolerate fluctuations between deploying a model that is typically computationally intensively dense and memory intensively very sparse. Because there are so many potential patterns and training flows, sparse computations require the flexibility, programmability, and efficiency of next-generation hardware instead of just adding Tera-FLOPs and memory bandwidth to meet the computational demands of machine learning. A good implementation of light methods on a friendly architecture can effectively assist in overcoming present barriers like enormous power, high machine costs, and lengthy training times. 

Numerous computational frameworks have been proposed in response to the growth of machine learning and artificial intelligence applications and their inherent properties. In addition to conventional CPU-based architectures, some examples are Google TPU, NVIDIA A100 Nvidia, Cerebras CS-2, Graphcore IPU, and SambaNova RDU. The entire extent of these hardware and software systems’ capabilities, particularly in handling a broad spectrum of sparse and dense applications, remains to be discovered, despite a few attempts to assess and compare these systems. Additionally, many of these frameworks are still privately owned and not accessible for public research in the public domain. Although promising, sparse approaches have additional difficulties besides architectural compatibility. 

The accuracy of a particular model, as opposed to a dense-only baseline, depends on a wide range of factors, including structured, semi-structured unstructured sparsity, percentages of sparsity weights/activation sparsity, and training schedule. These decision factors must be determined to get the most up-to-date metrics on a particular model, which takes time and effort. Large language models, which may accommodate a range of language applications, are widespread foundation models in the NLP sector, such as the 13B parameter GPT. Researchers from SambaNova Systems in this study use this model to demonstrate how sparsity may be successfully included in an end-to-end training cycle to attain equivalent accuracy metrics. 

They contribute in the following significant ways: 

• A thorough examination of how sparsity, fusion, and dataflow capabilities interact. 

• A demonstration of speedups over A100 using sparse GPT 13B on SambaNova RDU. 

• Analysis of the sparse 13B GPT model’s loss, zero-shot, and few-shot statistics in comparison to its dense baseline 

The paper itself has more details on their analysis. 


Check out the Paper. Don’t forget to join our 18k+ ML SubRedditDiscord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more. If you have any questions regarding the above article or if we missed anything, feel free to email us at Asif@marktechpost.com

🚀 Check Out 100’s AI Tools in AI Tools Club

Aneesh Tickoo is a consulting intern at MarktechPost. He is currently pursuing his undergraduate degree in Data Science and Artificial Intelligence from the Indian Institute of Technology(IIT), Bhilai. He spends most of his time working on projects aimed at harnessing the power of machine learning. His research interest is image processing and is passionate about building solutions around it. He loves to connect with people and collaborate on interesting projects.

🚀 [FREE AI WEBINAR] 'Optimise Your Custom Embedding Space: How to find the right embedding model for YOUR data.' (July 18, 2024) [Promoted]