Request to Add Attention Kernel to MIOpen

[zjchen77]

Attention mechanisms are widely used in deep learning models, particularly in large language models. And a flexible attention kernel can help users to build accelerated language models conveniently on AMD platform. I have already designed an attention kernel for variable text lengths. I believe that adding an attention kernel to MIOpen would be a valuable enhancement for the community. Thank you for considering this request.

[junliume]

@zjchen77 thanks! Our CK kernels https://github.com/ROCmSoftwarePlatform/Composable_kernel already supports and in the process of refining them, Indeed exposing these kernels via MIOpen might be useful in many cases.

[CAHEK7]

@JehandadKhan is this issue relater to multi-head attention as well?

[JehandadKhan]

@CAHEK7 That is correct.

[ppanchad-amd]

@junliume Has this request been added to MIOpen? Thanks!