Introduce WASI-NN for Machine Learning inference

[Ángel M]

Is your feature request related to a problem? Please describe.

WASI-NN is a standard for inferencing in WebAssembly. From the README:

wasi-nn is a WASI API for performing ML inference. Its name derives from the fact that ML models are also known as neural networks (nn). ML models are typically trained using a large data set, resulting in one or more files that describe the model's weights. The model is then used to compute an "inference," e.g., the probabilities of classifying an image as a set of tags. This API is concerned initially with inference, not training.

The goal is to provide AI / ML capabilities to Wasm Workers Server opening new use cases. This feature will enable developers to add ML capabilities to their applications. Currently, the models are loaded from the Wasm module using the Graph primitive. Then, developers must convert the data to the Tensor primitive, send it to a ML backend at the host and retrieve the output.

We will include an example so you can try and learn.

Describe the solution you'd like

Configuration

Following the capability-based security approach of wws, workers won't be able to call WASI-NN APIs by default. You will need to configure it via config file:

name = "nn"
version = "1"

[features]
[features.wasi_nn]
allowed_backends = ["openvino"]

Language support

For the initial release, WASI-NN bindings will be available only to Rust through the wasi-nn crate. In the future, we will add support for more languages like JavaScript and Python.

Machine Learning Backends

The initial host implementation will rely on the wasmtime-wasi-nn crate. This crate only supports OpenVINO™ as machine learning backend. However, in the future we will add new backends using different approaches.

Tasks

• Add WASI-NN bindings
• Add configuration properties
• Add example
• Add documentation