Revert "Update stable FA3 API (flash_api_stable.cpp) to be inline with the unstable ABI port (flash_api.cpp)"

[LucasWilkinson]

Reverts #152

This broke FLASH_ATTN_MLA_SPARSE (vllm-project/vllm#48221) not to mention changes a bunch of heuristics

[MatthewBonanni]

LGTM

[Harry-Chen]

not to mention changes a bunch of heuristics

@cleonard530 Just to check, did you notice this when drafting your PR? I personally glanced over the changes and did not notice this kind of changes

[cleonard530]

not to mention changes a bunch of heuristics

@cleonard530 Just to check, did you notice this when drafting your PR? I personally glanced over the changes and did not notice this kind of changes

I just checked this again. There are definitely a lot of heuristic changes between versions of flash_api_stable.cpp, but they were intentional to align flash_api_stable.cpp with flash_api.cpp (the diff between flash_api.cpp and flash_api_stable.cpp can be viewed here 7da1d8c).

@LucasWilkinson, are these the changes you're talking about between versions of flash_api_stable.cpp? If so, how should we proceed with migrating to the stable ABI? We can align flash_api_stable.cpp with flash_api.cpp first, before we change the build, and do more checks and testing that may not be covered by the CI before updating the CMake to use flash_api_stable.

[janeyx99]

Oh no this is bad :( But I'm a bit confused, as a pretty important key of the migration is that we shouldn't change functionality. @cleonard530 could you point out where the heuristics are different? From your diff and what you linked it looks like flash_api_stable and flash_api should be aligned with each other? Where do we deviate from the existing vllm behavior?

[janeyx99]

@LucasWilkinson I don't think this revert is warranted--I put some context in vllm-project/vllm#48221 (comment), can you confirm and let me know if there's anything I missed?