The tutorial is delivered as two 90-minute blocks. The morning block establishes the programming model, the GPU environment, and the first complete hybrid workflow; the afternoon block scales three further model families and the simulation strategy each one demands.
Focus: Establishing the CUDA-Q programming model, the GPU environment, and an end-to-end hybrid QML training loop.
Key Topics:
Why simulation throughput and validation govern which QML experiments are feasible at all, and how CUDA-Q and cuQuantum fit together.
GPU environment setup on NVIDIA Brev — no local installation or dependency conflicts.
Programming quantum kernels: qubits, gates, circuits, measurement, and noise modeling on the GPU nvidia target.
Learned Optimization for Variational Models: Meta-learning an optimizer for QAOA that generalizes across problem instances, with CUDA-Q and PyTorch in a single training pipeline.
Lunch
Focus: Three QML model families at scale, and the simulation technique each one requires.
Key Topics:
Quantum Fast Weight Programmers (QFWP): A parameterized quantum circuit that reprograms the weights of a classical network, giving sequence modeling without explicit recurrence.
Quantum-Inspired Kolmogorov-Arnold Networks (QKAN-LLM): Data re-uploading activations as parameter-efficient replacements for MLP blocks, from function fitting through GPT-2-scale transformer blocks on the cuTensorNet solver path.
Quantum-Enhanced SVM: A classical SVM with a quantum feature-map kernel, captured once as a tensor network and contracted in batch via cuTensorNet, with a multi-stream cuTensor backend for further HPC scaling.
Copyright © TUT-519, 2026.