look into ONNX runtime

[tharvik]

ONNX has a JavaScript API, it seems to be the new framework on the bloc.

currently, we are using TensorFlowJS everywhere. to support more runtime, we would need to first abstract from it.

• some of our types are typedef of tf's ones: Dataset, Weights
• aggregation, validator, data_loader, preprocessing are all using tf types

[JulienVig]

MNIST web training with ONNX

[martinjaggi]

roadmap (not very LLM specific but good)

transformers.js uses it, and would easily give us tokenizers and model converters to/from HF which is crucial

[bhack]

https://opensource.microsoft.com/blog/2024/02/06/on-device-training-training-a-model-in-browser/

[tharvik]

bad news, upstream removed training from their web package.

[JulienVig]

😭

[bhack]

Probably one way to gain momentum is to push for training in the new W3C forming standards:
https://github.com/webmachinelearning/webnn/blob/main/explainer.md?plain=1#L187

[tharvik]

per discussion with @martinjaggi and upstream, we can maintain our own fork with training support.