LFQAP2: Large‐Scale Quantum Artificial Intelligence Training Platform Based on Supercomputing
Open Access
- 1 January 2025
- journal article
- Published by Wiley in Quantum Engineering
- Vol. 2025 (1)
Abstract
Quantum artificial intelligence (QAI) has made significant progress in recent years. However, the lack of platforms for large‐scale QAI models training has constrained research on related algorithms, such as quantum large language models. In this paper, we introduce LFQAP2, a supercomputing‐based platform designed for training large‐scale QAI models. The platform employs a hybrid parallelization strategy that integrates data parallelism and model parallelism. For data parallelism, different threads compute the loss for distinct sample batches. For model parallelism, different processes compute the gradients for separate parameter batches. This parallelization strategy integrates the strengths of both data parallelism and model parallelism. It has high node‐internal computing efficiency and excellent scalability, and can reduce internode communication overhead, limiting it to the exchange of only a small number of floating‐point values. LFQAP2 is implemented using the MPI + OpenMP parallel programming model and deployed on a 16‐node cluster, where each node is equipped with 32 computing cores. To assess the platform’s efficiency, we evaluate its performance on two types of tasks: image recognition (MNIST dataset classification) and natural language processing (word embedding). For the 16‐qubit MNIST classification task with 128 trainable parameters, the 32‐thread acceleration efficiency reaches 0.852, while the 16‐node multiprocess acceleration efficiency is 0.357. For the 9‐qubit word embedding task with 288 trainable parameters, the 32‐thread acceleration efficiency also reaches 0.843, and the 16‐node multiprocess acceleration efficiency is 0.598. We successfully completed the training of a 288‐parameter model within 23 min, providing a robust computational platform to support research on large‐scale quantum models.Keywords
Funding Information
- National Natural Science Foundation of China (62472072)
This publication has 20 references indexed in Scilit:
- Quantum self-attention neural networks for text classificationScience China Information Sciences, 2024
- Strong generalization in quantum neural networksQuantum Information Processing, 2023
- Practical advantage of quantum machine learning in ghost imagingCommunications Physics, 2023
- Generalization in quantum machine learning from few training dataNature Communications, 2022
- Quantum generative adversarial networks with multiple superconducting qubitsnpj Quantum Information, 2021
- Simulation of quantum computing on classical supercomputers with tensor-network edge cuttingPhysical Review A, 2021
- A rigorous and robust quantum speed-up in supervised machine learningNature Physics, 2021
- Massively parallel quantum computer simulator, eleven years laterComputer Physics Communications, 2019
- Evaluating analytic gradients on quantum hardwarePhysical Review A, 2019
- Quantum Generative Adversarial LearningPhysical Review Letters, 2018