Dispatching circuit fragments in waves
Circuit cutting splits a quantum circuit that is too large for one device into many small fragments. Each fragment can run anywhere: on the centre’s own QPU, on a remote QPU behind a cloud queue, or on a classical simulator. DQR treats fragments as ordinary schedulable jobs, sends them out in waves without waiting on each one, and moves a fragment to a simulator when its QPU run fails, without restarting the rest.
This page runs a toy model with invented timings, to show the shape of the idea. Measured results on CESGA Qmio and IBM Torino are in the paper.
One fragment at a time
Submit to the local QPU, wait for the result, then submit the next. A failed run is retried on the same QPU.
Waves across every backend, with failover
Each wave goes out at once to whichever backend would finish it first. A failed QPU run moves to a simulator straight away.
- Local QPU run
- Remote QPU run
- Simulator run
- Waiting in the remote queue
- Failed run
- Wave boundary
What the model leaves out
Real fragments differ in width and depth, QPU calibration windows come and go, and reconstructing the original result from fragments has its own classical cost. The runtime in the paper also goes through the centre’s resource manager (Slurm) and keeps a backend-agnostic description of every fragment, which is what lets it choose and change backends per fragment.