youngkingdom/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix typo (#2136)

Browse Source 
Co-authored-by: XiaoDong <xiaod@nvidia.com>
This commit is contained in:
dongxiao
2025-03-18 10:19:43 +08:00
committed by GitHub
parent 6c6b78550e
commit bd03b22f64
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
media/docs/grouped_scheduler.md
View File

@ -340,7 +340,7 @@ Thus, there are 216 tiles across the group.
Suppose this grouped GEMM is run on GA100, which has 108 SMs. Suppose that
the occupancy given the parameters of the grouped GEMM is one -- one threadblock
can be active at a time on an SM. The grouped GEMM will, thus, run with 108
persistent threadblocks, each of which computes (256 / 108) = 2 tiles.
persistent threadblocks, each of which computes (216 / 108) = 2 tiles.
Under the round-robin assignment of tiles to threadblocks employed by
the grouped GEMM scheduler, the assignment of tiles to threadblocks