rocQuantum

rocQuantum is a clean-room, open-source implementation of NVIDIA's cuQuantum SDK for AMD GPUs running the ROCm/HIP stack. It provides production-grade primitives for high-performance quantum-circuit simulation, tensor-network contraction, open-quantum-system dynamics, Pauli-string propagation, and stabilizer formalisms on Instinct-class accelerators (MI100 / MI200 / MI300).

rocQuantum is not affiliated with or endorsed by NVIDIA. The cu* -> roc* API surface is reimplemented from scratch from publicly-documented behaviour. No proprietary headers or binaries are copied.

Component libraries

rocQuantum library	cuQuantum analogue	Purpose
librocstatevec	libcustatevec	Dense state-vector simulation (single- + multi-GPU)
libroctensornet	libcutensornet	Tensor-network construction, pathfinding, contraction
librocdensitymat	libcudensitymat	Density-matrix / open-quantum-systems dynamics
librocpauliprop	libcupauliprop	Heisenberg-picture Pauli-string propagation
librocstabilizer	libcustabilizer	Stabilizer-formalism evolution and QEC decoders

All five share a uniform handle/workspace lifecycle, async stream-based execution (hipStream_t), versioned headers under include/, and versioned shared-object sonames for ABI stability.

ROCm dependency map

rocQuantum builds on the existing ROCm math stack rather than reinventing dense linear algebra or collectives:

Purpose	rocQuantum uses	cuQuantum analogue
Runtime / device API	hip	cuda_runtime
Dense BLAS	rocBLAS, hipBLAS	cuBLAS
Dense solvers / SVD	rocSOLVER, hipSOLVER	cuSOLVER
Sparse BLAS	rocSPARSE	cuSPARSE
Tensor contraction	hipTENSOR	cuTENSOR
FFT	rocFFT, hipFFT	cuFFT
Random numbers	rocRAND	cuRAND
Primitives / sort / scan	rocPRIM, hipCUB	CUB
High-level templates	rocThrust	Thrust
Multi-GPU collectives	RCCL	NCCL
Runtime JIT	hipRTC	NVRTC

Supported GPUs

• Datacenter: gfx908 (MI100), gfx90a (MI200/MI250), gfx940/gfx941/gfx942 (MI300A/X)
• Workstation (dev only): gfx1100/gfx1101 (RDNA3)

Quick build

git clone https://github.com/ROCm/rocQuantum.git
cd rocQuantum
cmake -S . -B build \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_HIP_ARCHITECTURES="gfx90a;gfx942" \
  -DROCQUANTUM_BUILD_TESTS=ON \
  -DROCQUANTUM_BUILD_SAMPLES=ON
cmake --build build -j
sudo cmake --install build

The Python package is built independently:

cd python
pip install --no-build-isolation -e .

A reference Docker image is provided:

docker build -f docker/Dockerfile.rocm6 -t rocquantum:rocm6 .

Repository layout

rocQuantum/
  CMakeLists.txt            # superbuild driver
  cmake/                    # FindXxx finders + helpers
  libs/                     # the five native libraries
    rocstatevec/
    roctensornet/
    rocdensitymat/
    rocpauliprop/
    rocstabilizer/
  python/                   # Cython bindings + high-level Python API
  samples/                  # C/C++/HIP samples (one folder per library)
  benchmarks/               # roc-quantum-benchmarks Python package
  extra/                    # MPI comm-plugin, wheel-build helpers
  docs/                     # Sphinx documentation site
  docker/                   # ROCm 6 / ROCm 7 build containers
  ci/                       # self-hosted-runner job scripts

License

rocQuantum is released under the BSD 3-Clause license. See LICENSE. Each library publishes its own LICENSE file in its subdirectory; all are BSD-3-Clause unless explicitly noted.

This project is a clean-room reimplementation: no header, source, or binary from the proprietary NVIDIA cuQuantum SDK is incorporated. The API shape is reproduced based on the publicly-published documentation at https://docs.nvidia.com/cuda/cuquantum/latest/ for interoperability purposes only.

Citing rocQuantum

If you use rocQuantum in academic work, please cite us using the metadata in CITATION.cff. For a comparison benchmark, please additionally cite the upstream cuQuantum work by H. Bayraktar et al., QCE 2023, doi:10.1109/QCE57702.2023.00119.