Polyglot plugins

The same DAG runs with kernels in three languages. The repository ships the FFT + matrix-multiply pipeline three times — Rust (examples/matrix-compute, nalgebra + rustfft), C (examples/matrix-compute-C, FFTW + OpenBLAS), and Python (examples/matrix-compute-python, NumPy) — with identical topology (examples/README.md):

python examples/matrix-compute/run_bench.py --workers 4
python examples/matrix-compute-C/run_bench.py --workers 4
bash   examples/matrix-compute-python/run_bench.sh --workers 4

All three compute the same result. The C example's make validation target cross-checks its output against the Rust reference (examples/matrix-compute-C/README.md).

Why this works

The graph identifies kernels by function name only:

{ "name": "compute_fft", "factor": "num_nodes", "function": "compute_fft", ... }

The runtime never sees kernel source. It loads one plugin shared library, resolves each "function" string through the generated registry, and calls through a C ABI. What differs per language is only how the registry entry is produced:

Language	You write	Converter input
Rust	#[tomii_export] pub fn compute_fft(...)	the annotated .rs source
C	// @tomii_export(buffer: mut_array) above the prototype	the annotated header
Python	@tomii.export def compute_fft(v): ...	none — the bundled bridge plugin dispatches by name

This is the kernel half of tripartite decoupling (see What is Tomii): the graph, the kernels, and the runtime are separate artifacts, so you can swap the kernel language without touching the other two. No runtime source changes were needed for any of the three examples (repository README.md).

Mixing languages in one graph

Because nodes bind to registry names, a single graph can mix languages: any function the loaded plugin library exports under the expected name satisfies the node, regardless of what language produced it. The Python bridge is itself an example — a Rust dylib whose registered functions call into Python. The runtime loads one plugin library per run, so mixed-language kernels must be linked or bridged into that one library.

Choosing a language

• Rust: no system dependencies beyond cargo; the default for new plugins.
• C: reuse existing native libraries (FFTW, OpenBLAS, CUDA — see examples/gpu-vectoradd).
• Python: fastest iteration; parallel when kernels are NumPy/BLAS-bound or under free-threaded 3.13t (see Python plugins).

The examples page lists what each variant requires.