add training for LSTM models completely on device

[reidisaki]

🚀 The feature, motivation and pitch

I've seen lots of examples of tuning embeddings, centroids and adding weights to the dense heads of an LSTM model but I want to know if it's possible to basically pass in data for the model to be trained on fully on device.

This way there are no privacy concerns and it's all personalized on device. Currently I've seen use cases for LLM's and things that aren't personalized so you can keep the underlying model frozen but that won't fit my usecase.

Alternatives

With powerful devices coming out, and if we can have the model be trained on off hours when the users device is charging and plugged in I'm wondering if this is feasible, given that the data might not be large.

Additional context

No response

RFC (Optional)

No response

[GregoryComer]

We do have experimental support for training in ExecuTorch - see https://github.com/pytorch/executorch/tree/main/extension/training for details on usage. It has been intended for on-device fine tuning, but it should be possible to train a model entirely on device, given a sufficiently small model and long time scale.

[larryliu0820]

cc @JacobSzwejbka @lucylq

[lucylq]

We have an Android example of finetuning here as well: https://github.com/meta-pytorch/executorch-examples/tree/main/cifar/android/CifarETTrainingDemo