QFw

QFw is a quantum execution framework for running quantum applications against simulator or hardware services. It uses DEFw as the distributed runtime, adds QFw service APIs, and provides QPM services for execution targets such as TNQVM and NWQ-Sim.

A QPM is a Quantum Platform Manager. In QFw, a QPM service represents one execution target, advertises its type and capabilities, accepts circuit submissions through api_qpm_execution, and exposes separate admission, policy, scheduler, telemetry, and service-control bindings.

The same QFw application workflow can run on a local node, in a Slurm allocation, or inside the containerized QFw-SLURM-Cluster environment. The top-level scripts hide most of the differences between those launch modes.

Build QFw

Clone QFw with its submodules:

git clone --recursive git@github.com:openQSE/QFw.git
cd QFw

The submodules provide DEFw and the QHW Python/runtime packages used by QFw services. For an existing checkout that was not cloned with --recursive, initialize or update them before building:

git submodule update --init --recursive

Create or activate the shared Python environment that QFw should use:

python3 -m venv /path/to/qfw-venv
source /path/to/qfw-venv/bin/activate
python -m pip install --upgrade pip setuptools wheel
python -m pip install -r setup/build-requirements.txt
python -m pip install -r setup/requirements.txt

Install QFw with the CMake-backed installer. Use --with-defw when this install prefix should also build and install the bundled DEFw tree:

./setup/qfw_install.sh --prefix /path/to/qfw-install --with-defw

The same flow can be run manually:

cmake -S . -B build -DCMAKE_INSTALL_PREFIX=/path/to/qfw-install \
  -DQFW_BUILD_BUNDLED_DEFW=ON
cmake --build build
cmake --install build

The install places public commands in /path/to/qfw-install/bin, including qfw-activate, defw-python, qfw-setup, qfw-srun, qfw-teardown, qfw-dirsvc-start, and qfw-service-start.

Simulator runners and optional hardware client libraries must be available through the activated environment or site image. The QFw install packages the QFw service APIs, service modules, examples, QHW submodules, and runtime configuration templates; it does not build TNQVM, NWQ-Sim, QRMI, or QDMI-on-IQM from source.

Run QFw Locally

Activate QFw and run one of the example wrappers. When QFw is not inside a Slurm allocation, the launcher uses the local node for both the application and the services.

source /path/to/qfw-install/bin/qfw-activate
cd "$QFW_SHARE_DIR/examples"
./qfw_mpi_smoke.sh
qfw_deactivate

The example wrappers call qfw-setup, run one application through qfw-srun, and then call qfw-teardown. Do not call qfw_deactivate until the wrapper completes.

Run Examples

Activate QFw first, then run examples from the installed examples directory:

source /path/to/qfw-install/bin/qfw-activate
cd "$QFW_SHARE_DIR/examples"

Validate framework startup and Qiskit backend construction:

./qfw_init_test.sh

Run the MPI smoke service and verify an MPI payload can launch:

./qfw_mpi_smoke.sh

Run a simple Qiskit circuit through the NWQ-Sim backend:

./qfw_qiskit_simple.sh 4

Run GHZ through Qiskit:

./qfw_ghz.sh qiskit 4 nwqsim 4

Run GHZ through PennyLane:

./qfw_ghz.sh pennylane 4 nwqsim 4

Run the fixed PennyLane remote-backend example:

./qfw_pennylane.sh

Run the Qiskit QAOA Max-Cut example:

./qfw_qaoa.sh nwqsim
./qfw_qaoa.sh tnqvm

Run the Qiskit VQE example. The argument is the optimizer iteration limit:

./qfw_qiskit_vqe.sh 1

Run a SupermarQ workflow:

./qfw_supermarq.sh sync 1 4 128 false ghz nwqsim

Run a chemistry application script when the chemistry application tree is available:

./qfw_chem_app.sh <script-name.py>

Run the standard example set sequentially:

./qfw_run_all.sh

Useful qfw_run_all.sh overrides:

QFW_RUN_ALL_BACKEND=nwqsim ./qfw_run_all.sh
QFW_RUN_ALL_QUBITS=4 QFW_RUN_ALL_VQE_ITERS=1 ./qfw_run_all.sh

Examples that need a managed reservation should be launched through qfw_slurm_driver.sh. This script is the test stand-in for the future Slurm/SPANK integration: it reserves capacity, exports only QFW_RESERVATION_ID to the application step, runs the application through qfw-srun, and releases the reservation afterward.

The driver request carries the standardized reservation shape used by QPM and qhw-admission: target device, workload kind, walltime, qtask count, qubits, depth, one-qubit gate count, two-qubit gate count, shots, and measurement count. It also carries trusted launcher context such as user, job/allocation, and scope. Hardware runs may add a non-secret credential hint or opaque credential handle; provider API keys stay in the site-owned QPM environment and are not exported to the application.

./qfw_slurm_driver.sh \
  --backend nwqsim \
  --example supermarq \
  --qubits 4 \
  --depth 8 \
  --one-q-gates 4 \
  --two-q-gates 3 \
  --shots 128 \
  --count 1 \
  --workload-kind quantum \
  --operation async_run \
  --analytics-json '{"application":"supermarq"}' \
  -- ./qfw_supermarq.sh sync 1 4 128 false ghz nwqsim

qfw_supermarq.batch is a Frontier-oriented batch template. Edit the account, node counts, paths, and arguments before submitting it with sbatch:

sbatch qfw_supermarq.batch

For per-wrapper argument details, see examples/README.md.

Run On A Real Cluster

Cluster workflow

Use the same installed QFw prefix and example commands from the earlier sections. The cluster-specific work is taking the right allocation and selecting the site/runtime configuration used by qfw-setup.

On Frontier-style systems this is usually a two-component heterogeneous allocation:

salloc -N 1 -t 4:00:00 -A <project> --network=single_node_vni: \
  -N 1 -t 4:00:00 -A <project> --network=single_node_vni

Then source qfw-activate from the installed prefix and run the wrappers from Run Examples. The example commands are the same as the local workflow. Site defaults are read from $QFW_SITE_CONFIG, and per-job runtime overrides can be passed to qfw-setup by custom wrappers with --runtime-config.

Run With QFw-SLURM-Cluster

Containerized Slurm workflow

Use the QFw-SLURM-Cluster README for the container build, image pull, shared directory setup, and cluster startup steps.

Inside the container, QFw normally lives at:

/workspace/qfw-container-base/QFw

The QFw-SLURM-Cluster image already contains the simulator runners, so a development checkout normally only needs a QFw install prefix:

cd /workspace/qfw-container-base/QFw
./setup/qfw_install.sh --prefix /workspace/qfw-container-base/qfw-install-dev \
  --with-defw
source /workspace/qfw-container-base/qfw-install-dev/bin/qfw-activate

After activation, run the wrappers from Run Examples. The example commands are the same as the local workflow.

Install Configuration Reference

Installed layout and runtime configuration

QFw installation is CMake-backed. setup/qfw_install.sh is a convenience wrapper around:

cmake -S . -B build -DCMAKE_INSTALL_PREFIX=<prefix>
cmake --build build
cmake --install build

Important CMake options:

CMAKE_INSTALL_PREFIX: install prefix for QFw commands, Python modules, service APIs, examples, and configuration templates.
QFW_BUILD_BUNDLED_DEFW: build and install the bundled DEFw tree.
QFW_DEFAULT_DEFW_PREFIX: default DEFw prefix encoded into qfw-activate; use self when DEFw is installed into the same prefix.
QFW_PYTHON_INSTALL_DIR: relative Python site-packages destination.

The installed prefix contains:

bin/qfw-activate: prepares QFw, DEFw, Python, and library paths.
bin/defw-python: runs Python through the DEFw executor bridge.
bin/qfw-setup: creates runtime state and starts allocation-local services when requested.
bin/qfw-srun: launches applications with the active QFw runtime state.
bin/qfw-teardown: stops QFw-owned runtime processes and removes run state.
bin/qfw-dirsvc-start: starts a long-running or allocation-local DEFw directory service.
bin/qfw-service-start: starts a long-running or allocation-local QFw service.
share/qfw/config: site, runtime, service, and device configuration templates.
share/qfw/examples: installed example wrappers and application tests.

The default runtime configuration files are:

share/qfw/config/site.yaml: site directory and install-prefix defaults.
share/qfw/config/runtime.yaml: site-only resolver defaults.
share/qfw/config/runtime/local.yaml: allocation-local directory and QPM startup.
share/qfw/config/runtime/hybrid.yaml: allocation-local startup with site fallback.
share/qfw/config/services/local-services.yaml: allocation-owned simulator services, placement, and provider launch settings.
share/qfw/config/services/site-services.yaml: operator-started hardware QPM implementations.

qfw-activate exports QFW_PREFIX, QFW_SHARE_DIR, QFW_SITE_CONFIG, QFW_RUN_BASE_DIR, DEFW_PREFIX, DEFW_EXTERNAL_SERVICES_PATH, and DEFW_EXTERNAL_SERVICE_APIS_PATH. It can also activate a shared Python virtual environment before layering QFw paths:

source /path/to/qfw-install/bin/qfw-activate --venv /path/to/shared-venv

Without --venv, activation still works with the default installed environment. If a user virtual environment is already active, it is preserved and its site-packages are prepended to PYTHONPATH so services started through defw-python see the same Python package set as the application. When --venv is provided and a different virtual environment is already active, the explicit --venv path takes precedence and qfw-activate switches to it.

Simulator and hardware-provider dependencies are site/image concerns. Install them into the shared user environment or make them available through the site module stack before sourcing qfw-activate.

The IQM service does not return _raw_iqm as a separate result field. Instead, it embeds the full native IQM result payload in qhw_result["raw"] by default. To turn that off, set this before starting the IQM service:

export QFW_IQM_INCLUDE_RAW_RESULT=false

With that setting, QFw still returns qhw_result, including qhw_result["extensions"]["iqm.v1"], but omits the full raw IQM payload.

Shared Filesystem Behavior

Node-local and shared directory expectations

QFw does not enforce a shared filesystem across nodes in a Slurm allocation. The setup layer treats the run id as global state, propagates it to remote setup commands, and creates the required QFw temp directories on each node before writing logs or startup artifacts there.

This means the QFw infrastructure can operate with node-local temp directories for startup logs, service logs, pid files, and the PRTE DVM URI as long as the producer and consumer of each file run on the same node. In the current startup model, the directory service, QPM services, and DVM startup run on the group-1 head node, so those local files do not require a cluster-wide shared directory.

Backend simulators can have stricter requirements. QFw does not rewrite or stage simulator-specific files automatically.

QASM input files are written by QFw on the service node. This is safe for backends that only read the QASM file from rank 0 on that same node.

NWQ-Sim count output is rank-0 guarded and does not require every MPI rank to write a shared output file.

NWQ-Sim statevector dumps use NWQ-Sim's native --dump_file path. In the MPI statevector backend, every MPI rank participates in writing the same dump file in rank order. That requires the dump path to be visible and writable from all MPI ranks, unless the run is single-rank or head-local.

For node-local filesystems, avoid multi-node NWQ-Sim statevector dumps unless the dump directory is placed on shared storage or explicit staging is added. QFw can run without a shared filesystem, but individual simulator modes may still require one.

Developer Testing

Local CI-style checks

Run these checks when editing QFw itself:

python -m pip install flake8 pytest
./.github/scripts/ci-syntax.sh
./.github/scripts/ci-mock.sh

The example wrappers under examples/ are integration paths. They expect a configured and activated QFw environment.

High Level Design

Runtime architecture

QFw uses DEFw as the distributed runtime and layers QFw-specific services and APIs on top. DEFw handles process startup, messaging, role management, and remote execution. QFw adds simulator-specific QPM services, QRC execution paths, installation helpers, and example applications.

The repository is organized around:

setup/: the CMake install helper, runtime command implementation, and default service manifest.
services/: QFw-owned DEFw services such as svc_tnqvm_qpm and svc_nwqsim_qpm.
service-apis/: independent QFw-owned DEFw service APIs for QPM execution, admission, policy, scheduler, telemetry, and privileged control.
DEFw/: the distributed runtime submodule.
bin/: simulator runner binaries copied from dependency builds.
examples/: runnable examples and integration-style tests.

The Slurm runtime model uses group 0 for applications and group 1 for services and simulator execution. Local mode maps those roles onto the same node.

flowchart LR
    subgraph G0["Application role"]
        App["Application process"]
        API["QFw client API"]
    end

    subgraph G1["Service and execution role"]
        DirSvc["DEFw directory service"]
        DVM["PRTE DVM"]
        QPM["QPM services"]
        MPI["MPI launch path"]
        Target["Simulator or hardware"]
    end

    App --> API
    API <--> DirSvc
    API <--> QPM
    DirSvc <--> QPM
    QPM --> MPI
    MPI --> DVM
    MPI --> Target
    QPM -. direct service path .-> Target

The local service manifest, share/qfw/config/services/local-services.yaml, describes allocation-owned simulators and their provider launch settings. The site service manifest, share/qfw/config/services/site-services.yaml, describes operator-started hardware QPM implementations. site.yaml selects the active site manifest, device-access configuration, and common QPM retention settings.

QFw-specific services and APIs are loaded into DEFw through:

DEFW_EXTERNAL_SERVICES_PATH
DEFW_EXTERNAL_SERVICE_APIS_PATH

This keeps DEFw generic while allowing QFw services to evolve independently.

Service Statevector Contract

Common statevector payload

QPM services may expose simulator statevectors when a backend requests QFW_CAP_STATEVECTOR. Each service owns parsing of its simulator-native output, but services should publish statevectors through the common builder in services/util/qpm/statevector.py.

The common service payload is:

{
    "type": "statevector",
    "format": "complex128",
    "num_qubits": 4,
    "num_amplitudes": 16,
    "data": [[real, imag], ...],
    "source": "nwqsim",
}

Services should return this structure under the statevector key:

{
    "counts": counts,
    "statevector": statevector.to_dict(),
}

Count-only services may continue to return the existing plain counts dictionary. The QFw backend accepts both forms.

NWQ-Sim writes statevectors through its native --dump_file option when QFw requests statevector data. QFw names the dump file from the circuit UUID with a .dump extension, parses it as the NWQ-Sim native complex128 statevector, converts it to the common payload, and removes the dump file after parsing or failure cleanup.

For MPI statevector runs, NWQ-Sim expects that dump path to be visible to all MPI ranks. Use shared storage for the dump directory or keep the run single-rank/head-local when running on node-local filesystems.

Navigation Menu

Repository files navigation

QFw

Table Of Contents

Build QFw

Run QFw Locally

Run Examples

Run On A Real Cluster

Run With QFw-SLURM-Cluster

Install Configuration Reference

Shared Filesystem Behavior

Developer Testing

High Level Design

Service Statevector Contract

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
.github		.github
DEFw @ 7728f89		DEFw @ 7728f89
backends		backends
bin		bin
cmake		cmake
docs		docs
examples		examples
external		external
patches/tnqvm		patches/tnqvm
service-apis		service-apis
services		services
setup		setup
share/qfw		share/qfw
tests		tests
.gitmodules		.gitmodules
CMakeLists.txt		CMakeLists.txt
LICENSE		LICENSE
README.md		README.md

Latest commit

History

Folders and files

Repository files navigation

QFw

Table Of Contents

Build QFw

Run QFw Locally

Run Examples

Run On A Real Cluster

Run With QFw-SLURM-Cluster

Install Configuration Reference

Shared Filesystem Behavior

Developer Testing

High Level Design

Service Statevector Contract

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages