[Issue]: FMHA change of drop_seed_offset to std::variant is breaking builds

[iratebadger]

Problem Description

The change of drop_seed_offset from a std::pair<uint64_t, uint64_t> to a std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>> is breaking builds in a trivial manner. This should have become 2 or more constructors one with each of the types that is then passed to a std::variant version explicitly. As it stands it requires some kind o ugly code for end users.

Operating System

24.04

CPU

Ryzen 7 / NA

GPU

AMD Instinct MI100

Other

No response

ROCm Version

ROCm 6.0.0

ROCm Component

Composable Kernel

Steps to Reproduce

Pull latest xformers or flash_attention and try to build using latest CK

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

This is a fixable issue but it requires all end users to change their code and I think there is a fix here that could result in better outcomes.

[schung-amd]

Hi @iratebadger, slight clarification: it looks to me like prior to this change the input was an std::tuple, not std::pair, is that correct? I think a pair should be implicitly converted to the variant, but the tuple would not and I can take a look at adding back in support for that.