torch._VF frontend functions

[Sam-Armstrong]

Frontend functions need to be added for the following torch._VF functions:

• torch._VF.rnn_tanh
• torch._VF.rnn_relu
• torch._VF.gru
• torch._VF.rnn_tanh_cell
• torch._VF.rnn_relu_cell
• torch._VF.lstm_cell
• torch._VF.gru_cell

[muzakkirhussain011]

Hi @Sam-Armstrong ,
Are there any specific considerations or challenges we should be aware of when adding frontend functions for these torch._VF operations?

• torch._VF.rnn_tanh
• torch._VF.rnn_relu
• torch._VF.gru
• torch._VF.rnn_tanh_cell
• torch._VF.rnn_relu_cell
• torch._VF.lstm_cell
• torch._VF.gru_cell

I'm wondering if there are existing patterns or best practices we should follow to ensure consistency and performance. Any thoughts on how we should approach this?

[Sam-Armstrong]

@muzakkirhussain011 for the gru, lstm_cell and gru_cell implementations, you can generally follow how torch._VF.lstm was written. So the call graph would look something like torch_frontend._VF.lstm -> torch_frontend.lstm -> ivy.lstm.

If we can, the optimal implementation for torch._VF.gru would be something like:

torch_frontend._VF.gru -> torch_frontend.gru -> ivy.gru -> backend.gru

where you implement ivy.gru for all backend frameworks (torch, tensorflow, jax and numpy - but this isn't strictly necessary - you can just use an ivy compositional implementation as has been done in the case of lstm. Just note you'll have to add torch_frontend.gru, as this also doesn't exist in the torch frontend apis yet.

Feel free to message me on discord if you're unsure about anything.