Transformers.js V4: Native WebGPU EP, repo restructuring, and more!

[xenova]

trafficstars

This is the official, long-awaited PR that introduces Transformers.js V4. Although it's currently still in draft mode, I'll be posting updates here for early review!

• [x] Migrate to and test new Native WebGPU EP.
  • Closes https://github.com/huggingface/transformers.js/issues/1435
  • Closes https://github.com/huggingface/transformers.js/issues/1424
  • Closes https://github.com/huggingface/transformers.js/issues/1418
  • Closes https://github.com/huggingface/transformers.js/issues/1416
  • Closes https://github.com/huggingface/transformers.js/issues/1320
  • Closes https://github.com/huggingface/transformers.js/issues/1309
  • Closes https://github.com/huggingface/transformers.js/issues/1307
  • Closes https://github.com/huggingface/transformers.js/issues/1308
  • Closes https://github.com/huggingface/transformers.js/issues/1217
• [x] Upgrade ONNX Runtime versions
  • https://github.com/huggingface/transformers.js/issues/1403
• [x] Move examples out of main transformers.js repo and into examples repo
• [x] API fully backwards compatible with v3
• [x] Migrate certain ONNX models on the HF hub to be v3 & v4 compatible (mainly focused on older models).
• [ ] Auto-device detection and improved defaults, depending on the end-user’s resources.
• [ ] Create an official Transformers.js Playground. This involves curating a list of recommended models for each task, as well as an interactive web-app/playground for users to explore what the library has to offer. We'd also like to add a simple copy-paste option for sample code so that users can get started much easier!
• [ ] Upgrade default pipeline models
• [ ] Migrate Tokenizer-specific logic (tokenizers.js and tests) to separate @huggingface/tokenizers library.
• [ ] Complete conversion script overhaul. Introduces significant performance optimizations.
  • Closes https://github.com/huggingface/transformers.js/issues/1415
  • Closes https://github.com/huggingface/transformers.js/issues/1402

See benchmarks

https://huggingface.co/onnx-community/all-MiniLM-L6-v2-ONNX:

https://huggingface.co/onnx-community/bge-base-en-v1.5-ONNX:

• [ ] Improved repository formatting & consistency
  • [ ] Organize modeling code into separate folders (./src/models/), grouped by model type -- models.js is getting pretty large!
  • [ ] Formatting will happen later to keep the number of changes between v3 and v4 lower.
• [ ] Documentation improvements and updates (particularly for older models).
• [ ] Fix typing issues
  • Closes https://github.com/huggingface/transformers.js/issues/1409
  • Closes https://github.com/huggingface/transformers.js/issues/1448
• [ ] React native support
• [ ] New models:
  • Chatterbox. Closes https://github.com/huggingface/transformers.js/issues/1434
  • Mamba/Mamba2 (e.g., Granite-4.0-h)