Installation
QECTOR Decoder v3 ships as a Rust-compiled Python wheel (v1.0.1: 18 binary wheels for CPython 3.9-3.14, no sdist). One command on Linux x86_64, macOS ARM64, and Windows amd64. CUDA batch decoding is optional; OpenCL is not in the wheels and requires a source build.
System Requirements
- CPython 3.9-3.14
- 64-bit operating system (Linux x86_64, macOS arm64, Windows amd64)
- Stim/Sinter/PyMatching are optional extras; install
qector-decoder-v3[stim]when needed. - GPU paths are optional and environment-dependent; follow the package release documentation for supported drivers and extras. OpenCL is not included in the published wheels: it requires a source build.
Quick Install (Recommended)
This installs the latest stable release with pre-compiled binary wheels.
Optional extras: [stim] (Stim/Sinter/PyMatching), [bench] (local measurement tools), [all] (full environment).
Verify Installation
Platform-Specific Notes
Linux
Use the current PyPI wheel and the release documentation for supported Linux platforms and runtime dependencies.
macOS
Use the current PyPI wheel and the release documentation for supported macOS architectures and minimum version.
Windows
Use the current PyPI wheel for supported Windows Python and architecture combinations. GPU availability remains driver- and device-dependent.
Optional GPU Workflows
GPU availability depends on the installed wheel, optional dependencies, operating system, driver, and device. Use the extras and support matrix documented by the package release rather than assuming a GPU backend is available.
Since 1.0.1 every paid licence (Pro or Enterprise) unlocks GPU batch decoding. The OpenCL backend is not part of the published wheels: OpenCLBatchDecoder requires a source build with --features opencl, and PyPI carries wheels only (no sdist).
Troubleshooting
ImportError on macOS: Ensure you have Python 3.9-3.14 from python.org or Homebrew. The system Python may not work.
GPU not detected: Verify CUDA installation with nvidia-smi. Check the package release documentation for supported CUDA versions.
Slow import: First import initializes the runtime and validates optional paths. Subsequent imports are fast.