Pyodide

cuNumPy uses Pyodide’s NumPy build with the existing numpy backend. No separate backend or special wheel is needed. The supported path is CPU array operations and ordinary Python callables wrapped in PyccelKernel. CuPy/CUDA execution and native Pyccel compilation are outside this support.

Installation and use

After initializing Pyodide, load micropip and install cuNumPy. Its dependency resolver supplies NumPy and array-api-compat; see Pyodide package loading.

await pyodide.loadPackage("micropip");
await pyodide.runPythonAsync(`
    import micropip
    await micropip.install("cunumpy")

    import cunumpy as xp
    xp.set_backend("numpy")

    def scale(values, factor):
        values[:] *= factor
        return values

    values = xp.array([1.0, 2.0, 3.0])
    result = xp.PyccelKernel(scale)(values, 2.0)
    assert result is values
    print(xp.to_numpy(result))  # [2. 4. 6.]
`);

NumPy is selected by default when ARRAY_BACKEND is unset. Leave it unset or set it to numpy before importing cuNumPy. to_numpy and to_cunumpy preserve existing NumPy arrays, including views. synchronize and set_device are no-ops on this backend.

PyccelKernel does not import Pyccel or compile functions. With NumPy arguments and its default settings, it calls your function directly: argument identity, in-place mutation, return values, and exceptions retain normal Python semantics. The outputs option controls GPU copy-back only; it does not prevent a Python kernel from mutating CPU arguments. Keep use_cupy unset or False in Pyodide.

Applications such as Struphy must choose Python-source kernels and ensure their imports (including any Pyccel decorators) work without a native compiler. Wrapping a function does not make its own dependencies Pyodide-compatible.

Testing the built wheel

From the repository root, with Python and Node.js 22 installed:

python -m pip install build '.[test]'
python -m build --wheel
python -m pytest -q tests/portable
npm ci --prefix tests/pyodide
node tests/pyodide/run.mjs dist/cunumpy-*-py3-none-any.whl

Pass exactly one wheel (an explicit path if dist contains multiple versions). Network access is required to download the Pyodide packages and wheel dependencies.

CI pins Pyodide 314.0.6 and executes the portable suite in its actual WebAssembly Python runtime under Node.js. The runner installs the built wheel with micropip, including dependency resolution, without exposing the source tree. It checks the default NumPy backend, installed version metadata, and rejects attempts to import CuPy or Pyccel during import and CPU execution. The suite covers array operations, conversions, backend context restoration after normal and exceptional exits, and Python-kernel return values, mutation, aliasing, and exceptions.

This is a runtime compatibility check; browser-specific page integration, workers, and hosting configuration are not exercised by this CI job.

The test extra installs compiler-free test dependencies. Native compiled-kernel tests use the separate test-compiled extra and require a working compiler.