Current quantum processors have a limited qubit capacity, preventing large combinatorial optimization problems from being solved directly on a single QPU. QSplit addresses this limitation from a hybrid HPC-quantum perspective, combining problem decomposition with heterogeneous execution to solve optimization problems beyond the capacity of individual quantum devices.
The QSplit framework
QSplit is a workflow-oriented framework for distributed hybrid quantum-classical optimization. It targets combinatorial optimization problems expressed in the Quadratic Unconstrained Binary Optimization (QUBO) form and follows a split-solve-aggregate approach. Large problem instances are first decomposed into smaller subproblems whose size is compatible with the available optimization backends. The resulting subproblems can then be solved independently using heterogeneous classical or quantum resources, and their partial solutions are finally aggregated to reconstruct a solution for the original problem.
This approach makes it possible to investigate optimization problems that exceed the direct capacity of current QPUs while also exposing parallelism across the generated subproblems. QSplit supports different optimization backends, including Simulated Annealing and QAOA, allowing the same decomposition strategy to be evaluated using both classical and quantum approaches.
A central aspect of QSplit is the relationship between problem structure and decomposition quality. The framework has been evaluated on Max-Cut and Knapsack instances with different structural characteristics. Experimental results show that decomposition is particularly effective for sparse problems, where enough of the original structure can be preserved for effective aggregation. Dense instances, instead, expose the limitations of purely structural decomposition, since splitting the original problem can remove global correlations that are relevant to the final solution.
The experiments demonstrate the possibility of addressing optimization problems up to 10 times larger than the available QPU qubit capacity, showing how decomposition and heterogeneous execution can extend the range of problems accessible to current quantum resources.
StreamFlow application
StreamFlow provides the orchestration layer that allows QSplit to combine heterogeneous computing environments within the same workflow. In particular, the same optimization pipeline can be configured to execute across classical HPC resources and remote quantum infrastructures, while keeping the high-level workflow logic separated from the specific execution environment. This makes it possible to experiment with different combinations of resources and solver backends without redesigning the overall application workflow.
The split, solve, and aggregate phases are represented as separate workflow stages, allowing StreamFlow to coordinate their execution and the movement of intermediate results across the resources involved.
QSplit therefore provides a concrete example of how a workflow management system can support hybrid HPC-quantum applications: rather than treating quantum execution as an isolated computation, quantum resources become components of a larger distributed workflow together with classical computing resources. StreamFlow manages this heterogeneous execution while QSplit focuses on the optimization strategy, providing a portable and extensible environment for investigating the current scalability boundaries of hybrid quantum-classical computing.
M. Bifulco, F. Medina, D. Medić, L. Roversi, and M. Aldinucci, “QSplit: A workflow-oriented hybrid quantum-classical optimization framework,” in Euro-Par 2026: Parallel Processing - 32nd European Conference on Parallel and Distributed Processing, Proceedings, Part II, vol. 16782, pp. 195–209, 2026. doi:10.1007/978-3-032-35251-4_14