PyTorch Support

[deevashwer]

I see that SPU already has experimental support for PyTorch. Has it been tested on a transformer model yet (even for inference)?

[anakinxc]

Hi @deevashwer

PyTorch support is still at it's very early stage, the demo is just showing it's potential...

Technically things like this might work fine with SPU.

[deevashwer]

Thanks! Is PyTorch support a priority for the SecretFlow team? If yes, is it possible to give a rough estimate on the timeline for this?

[anakinxc]

Thanks! Is PyTorch support a priority for the SecretFlow team? If yes, is it possible to give a rough estimate on the timeline for this?

Better PyTorch support is definitely something we would like to see, but it's not a priority at this point.

Feedbacks and inside thoughts that can help us make a better judgement are welcome :D

[deevashwer]

You probably already know this, but I would vouch for PyTorch simply because it is by far the most widely adopted framework and it unlocks the entire huggingface library. If you look here, everything is supported in PyTorch, and TF and JAX support is relatively sparse. All the state-of-the-art models (not just in NLP) are also typically released in PyTorch. So you stand to unlock all the models, if a robust support for PyTorch is added to SecretFlow.

[anakinxc]

You probably already know this, but I would vouch for PyTorch simply because it is by far the most widely adopted framework and it unlocks the entire huggingface library. If you look here, everything is supported in PyTorch, and TF and JAX support is relatively sparse. All the state-of-the-art models (not just in NLP) are also typically released in PyTorch. So you stand to unlock all the models, if a robust support for PyTorch is added to SecretFlow.

Thanks for the inputs, I'll bring this up to our team.

[github-actions]

Stale issue message. Please comment to remove stale tag. Otherwise this issue will be closed soon.

[anakinxc]

Hi @deevashwer

We just made some improvements on Torch support, check here. Feel free to give it a shot.

Be aware, due to an issue in huggingface/transformers, we do not support models from huggingface at this point.