LOGICAL-QAOA

Code for the paper Toward Fault-Tolerant Variational Optimization: QAOA under [[4,2,2]] Error Detection presented at Q-WORKS colocated with EuroPar 2026 at Pisa, Italy. A noisy QAOA framework under the [[4,2,2]] error detection code, optimized for large scale execution.

GPU Execution

Setup

1. Install Nvidia drivers

2. Install Nvidia Cuda

3. Install Nvidia cuQuantum

1. Install Docker
2. Install Nvidia Container Toolkit
3. Install Nvidia cuQuantum

Run Simulations of .json Circuits

1. Start the docker container in interactive mode. This also automatically activates a conda env with cirq

sudo docker run --gpus all -v "$(pwd)":/workspace -w /workspace -it --rm nvcr.io/nvidia/cuquantum-appliance:25.11-x86_64

cd logical-qaoa/

3. Run the python script

# If your setup supports NVLink (e.g. NVIDIA GH200)
python -m logical_qaoa.hpc_execution.hpc_execute --nlogical=4 --nshots=10000 --ngpus=4 --split --input circuit.json
# If you need to manually split shots among GPUs (e.g. NVIDIA L40S)
python -m logical_qaoa.hpc_execution.hpc_execute --nlogical=4 --nshots=400 --ngpus=4 --split --no-save --input circuit.json

[!note] The --nlogical argument corresponds to the number of logical qubits in your circuit. In the sqare max-cut qaoa it is 4.

If results are written to a file they can be found in logical_qaoa/hpc_execution/sim_results/.