Executorch - Encoder/Decoder models?

ExecuTorch
1

Not sure if this is the right place, but I am trying to understand what people would recommend for implementing an encoder-decoder architecture in executorch. Would you implement them as 2 separate models and call the encoder once in the c++ runner, then pass the results as an input to the decoder (followed by the same type of decoder iteration that decoder only models do?)

Or is there a better way?

Thanks

2

Would the alternative be to wrap both modules into a single parent module and execute them together? If so, I would not expect to see a difference between these approaches and would claim it comes down to your coding style and if you want to use the intermediates in another way.