Workshops Archive

QSPredictor: Runtime Prediction for Tensor-Network based Quantum Circuit Simulation on GPUs


Workshop: The 2nd International Workshop for Software Frameworks and Workload Management on Quantum-HPC Ecosystems

Authors: Chianing Wang and David Quiroga (Toyota InfoTech Lab, Toyota North America) and Younghyun Cho (Santa Clara University)

Abstract: Tensor networks enable the simulation of large quantum circuits with limited entanglement, potentially reaching hundreds of qubits. For circuits with higher entanglement, predicting runtime before execution is important for assessing simulation feasibility, wait times, and resource utilization in quantum-centric computing systems. Metrics that allow runtime estimation, such as treewidth, are NP-hard to calculate and change with each specific circuit, complicating usability in HPC schedulers. Circuits belonging to different families or algorithms present a similar structure that allows grouping through cheap characterization, providing information relevant to total runtime. We present QSPredictor, a Transformer-based predictor of tensor network planning, contraction, and total runtime using structural circuit properties and OpenQASM text, and compare it against a family-aware baseline. To improve data efficiency, QSPredictor selects representative training circuits followed by conditioned k-means++ selection. Under the complete proposed pipeline, QSPredictor achieves a total-runtime MAPE of 7.09% and a log-space R² of 0.9906.


Back to The 2nd International Workshop for Software Frameworks and Workload Management on Quantum-HPC Ecosystems Archive Listing

Back to Full Workshop Archive Listing