Bolt: An Automated Deep Learning Framework for Training and Deploying Large-Scale Search and Recommendation Models on Commodity CPU Hardware

By Nicholas Meisburger et al

Published on Oct. 21, 2023

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Related Work
3. Background: The SLIDE Algorithm
4. BOLT
4.1 Core Library
4.2 Automated Sparsity Hyperparameter Tuning
4.3 Sparse Inference

Summary

The document discusses the introduction of BOLT, a sparse deep learning library for training large-scale search and recommendation models on standard CPU hardware. It highlights the challenges associated with training massive models and the need for efficient, cost-effective solutions. BOLT aims to democratize deep learning capabilities by providing a high-level API for constructing models and automatically tuning specialized hyperparameters. The paper evaluates BOLT on various information retrieval tasks and showcases its competitive performance, cost-effectiveness, and faster inference time. BOLT has been successfully deployed in real-world scenarios, with a case study in the field of e-commerce provided. The document also explains the SLIDE algorithm, which uses Locality Sensitive Hashing to reduce computational requirements in deep learning. Additionally, BOLT features automated sparsity hyperparameter tuning and supports dynamic sparsity in inference to reduce latency.