pywebnn

Python bindings for the W3C WebNN API, powered by rustnn.

Install

pywebnn is available on PyPI:

pip install pywebnn

Development

Install the development dependencies and native extension in editable mode:

python -m pip install --upgrade pip
python -m pip install maturin numpy tokenizers
maturin develop

Re-run maturin develop after changing Rust code. To build with TensorRT RTX support, use maturin develop --features trtx-runtime and ensure the matching TensorRT runtime DLLs are available on PATH before launching Python.

Docs

Full documentation is published on GitHub Pages:

https://rustnn.github.io/pywebnn/

Quick Start

import numpy as np
import webnn

ml = webnn.ML()
context = ml.create_context(device_type="cpu")
builder = context.create_graph_builder()

x = builder.input("x", [2, 3], "float32")
y = builder.input("y", [2, 3], "float32")
out = builder.relu(builder.add(x, y))
graph = builder.build({"output": out})

result = context.compute(
    graph,
    {
        "x": np.array([[1, 2, 3], [4, 5, 6]], dtype=np.float32),
        "y": np.array([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], dtype=np.float32),
    },
)
print(result["output"])

TensorRT backend

TensorRT is opt-in. Build an editable installation with RustNN's TensorRT feature, then request it when creating the context:

maturin develop --features trtx-runtime

context = webnn.ML().create_context(backend="trtx")

This requires a compatible TensorRT installation and GPU. Use context.backend_info() to inspect the requested backend and compiled features. On Windows, make the matching TensorRT RTX runtime DLLs available through PATH before starting Python. The DLL version must match the TensorRT headers used when building the extension. Other opt-in runtime plugins are litert-runtime and cann-runtime; select them with backend="litert" and backend="cann", respectively.

Links

• GitHub: https://github.com/rustnn/pywebnn
• PyPI: https://pypi.org/project/pywebnn/
• Issues: https://github.com/rustnn/pywebnn/issues