Haicheng Wu comments

Results 323 comments of


                                            Haicheng Wu

[QST] Terminology question on GMMA::ScaleOut::One

The one in the mainloop is redundant, but it does not hurt anything. It has to be `One` in the mainloop.

Example77 blackwell FMHA decode attention optimization

@alihassanijr , could you please take a look? CC += @richardmcai

[BUG] Copy constructor of tma descriptor produces a corrupted copy

@thakkarV @ANIKET-SHIVAM

Add SM80/89 blockwise scaling kernel, support FP8 block/groupwise on Ada, INT8 on Ampere

@jackkosaian

[FEA] Has CUTLASS considered supporting Zero-points and block-wise scaling in Hoppr Mixed Grouped Gemm recently?

cc += @alexsamardzic @IwakuraRein

[BUG] Incorrect calculation results of CuteDsl/ampere/tensorop_gemm.py

@keithzzzzz

Improvements for: Groupwise scaling along M for FP8 gemm

@LucasWilkinson , we upstreamed our change to groupwise scaling kernels. there are some conflicts in this PR that needs to be solved. Our change is mainly: ``` Extend groupwise scaling...

Improvements for: Groupwise scaling along M for FP8 gemm

@manishucsd

Allow cluster sizes across m,n,k to be reported in cutlass profiler

@itramble , could you please review first?

[QST] link invalid in efficient_gemm.md

sorry, seems gtc pulled off the old ones.