Vectorized epilogue in v3 API

[masahi]

Hi, I've spent some time studying the v3 API. One of the things that's not clear to me is how vectorized epilogues would work in the v3 API. My understanding is that existing implementations in https://github.com/NVIDIA/cutlass/tree/master/include/cutlass/epilogue/collective only support scalar epilogue.

Since vectorized epilogues operate on the Array<T, N> type, do we need to "convert" cute::Tensor to Array<T, N>? Or will vectorized epilogue be natively supported on cute::Tensor slices in a future release?

[hwu36]

Stay tuned. It is coming

[thakkarV]

Hi! have you seen example 50 in the meantime?

[masahi]

Yes, I know it vectorizes copying but not math. And my target is sm80 for now.