RISE project reaches 1,500 Python packages with native RISC-V wheels
The RISE project has achieved 72% coverage of PyPI's most-downloaded native-code packages for riscv64, including major data science and ML libraries, dramatically reducing install times from source compilation.
Summary
The RISE project, led by Ludovic Henry at Qualcomm, has published prebuilt riscv64 wheels for 1,506 of the 2,076 most-downloaded PyPI packages that contain native code. This milestone covers critical scientific and machine learning stacks including NumPy, SciPy, pandas, cryptography, PyArrow, PyTorch and vLLM. The project tracks the top 15,000 most-downloaded packages on PyPI, filtering for those requiring compilation from C, C++, Rust or Fortran sources.
Prebuilt wheels eliminate lengthy source compilations: installing NumPy from a wheel takes approximately 25 seconds on a VisionFive 2 board versus roughly 15 minutes to build from source. The PyTorch 2.13.0 wheels were built natively on RISC-V hardware rather than through emulation, and passed 212,038 test cases in a full native test run. The project maintains automated CI through GitHub Actions, with builds on native hardware, clean-environment testing, and automated index regeneration. 519 packages in the tracked set still lack usable wheels, and the project's explicit goal is to upstream riscv64 wheel publishing so that RISE builds can eventually be retired.
Build infrastructure and package coverage
- Package scope and tracking: The project monitors the 15,000 most-downloaded PyPI packages, of which 2,076 contain native code requiring architecture-specific builds; 1,506 now have riscv64 wheels available through PyPI or the RISE package index.
- Performance impact: On a StarFive VisionFive 2 board, NumPy installation dropped from approximately 15 minutes of source compilation to roughly 25 seconds using a prebuilt wheel.
- Native build verification: PyTorch 2.13.0 wheels were compiled on physical RISC-V hardware rather than cross-compiled or emulated, with the full native test suite executing 212,038 test cases successfully.
- CI and publishing pipeline: Contributions require GitHub Actions workflows, updated package metadata, and versioned patches; the pipeline builds on native hardware, validates in clean test environments, publishes releases, and regenerates the package index automatically.
- Remaining gap: 519 tracked packages still lack usable riscv64 wheels, and the project's stated objective is to encourage upstream maintainers to publish official riscv64 wheels so RISE can sunset its own builds.
Why it matters
This development significantly lowers the barrier to Python-based data science, machine learning, and general software development on RISC-V Linux systems and single-board computers like the VisionFive 2. By providing prebuilt wheels for PyTorch and vLLM, the RISE project enables local AI inference and training workflows on open hardware without requiring users to maintain complex cross-compilation environments or wait hours for dependency resolution.
Original reference: blog.adafruit.com →