[QST] When to use MainloopSm90TmaGmmaWarpSpecializedFP8?

[ginowu]

Hi, there
In sm90_mma_tma_gmma_ss_warpspecialized_fp8.hpp, mma_promotion_interval=4 means will add current 4 MMA's result sum to the ultimate result. My question how could this non-default behavior could improve FP8 accuracy? And could you share some best practice on using this specialized implementation, like how to set mma_promotion_interval according to activation input range?

Thanks!

[ginowu]

By the way, all the collective mainloop specializations under include/cutlass/gemm/collective/ have the "mma_promotion_interval" member in Arguments, I understand this treatment makes uniform mainloop argument at host side possible. So the missing of mma_promotion_interval in below file is unexpected?

include/cutlass/gemm/collective/sm90_mma_array_tma_gmma_ss_warpspecialized.hpp

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[ginowu]

I've got the reason of this whole mma promotion by reading the deepseek v3 paper.