feat: exposing onnx backend to JS land

[kallebysantos]

What kind of change does this PR introduce?

Feature, Enhancement

What is the current behavior?

When using huggingface/transformers.js the model inference are wasm handled with onnxruntime-web library.

What is the new behavior?

This PR adds compatibility to exposed ORT feature. It allows to use transformers.js with the current native rust ORT and all its optimizations.

How to use it

The onnx backend is implicitly exposed to transformers.js v3 and the only code change needed from final users is to add device: 'auto'.

Currently there are an open PR to simplify it, but while it wasn't merged yet we need to follow the example bellow:

import { env, pipeline } from 'https://cdn.jsdelivr.net/npm/@huggingface/[email protected]';

// Ensure we do not use browser cache
env.useBrowserCache = false;
env.allowLocalModels = false;

const pipe = await pipeline('feature-extraction', 'supabase/gte-small', { device: 'auto' });

const output = await pipe("This embed will be generated from rust land", { pooling: 'mean', normalize: true });

---

Tested pipelines

NLP

• feature-extraction
• fill-mask
• text-classification
• token-classification
• zero-shot-classification
• text-generation
• text2text-generation
• translation
• summarization
• question-answering

Computer vision

• image-feature-extraction
• image-classification
• zero-shot-image-classification

This pipelines was only manually tested:

• object-detection *
• depth-estimation *
• image-segmentation *

[laktek]

@kallebysantos Looks like you'd need to rebase :)

[laktek]

LGTM! thanks for the contribution @kallebysantos 🎊

[github-actions]

🎉 This PR is included in version 1.62.0 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀