Description
Currently torch compiling the default llama model will generate a warning re: being unable to lower complex numbers.
torch/_inductor/lowering.py:1639: UserWarning: Torchinductor does not support code generation for complex operators. Performance may be worse than eager.
This is from the current rope embeddings impl.
This issue is to track it as it may be disconcerting to a user to see this warning.
1 - we could investigate suppressing the error if we confirm there is really no speed change.
2 - could look at non complex /compile friendly rope impl ala:
https://github.com/lessw2020/PyTorch_MultiModal/blob/2d0ab30e0f85c5e1e68d3ab62126a91e41188e82/torchmultimodal/modules/layers/position_embedding.py#L174
current plan is to review this after impl has stabilized at which point we can compare speed and loss curves.