NVIDIA/cutlass
1
0
Fork 0
mirror of https://github.com/NVIDIA/cutlass.git synced 2026-04-19 22:38:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix SFB Layout scale granularity representation (#2924)

Browse Source
This commit is contained in:
Aidan Do
2026-01-06 20:55:21 -08:00
committed by GitHub
parent 61b560983a
commit 670480df3a
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples/81_blackwell_gemm_blockwise/README.md
View File

@@ -19,7 +19,7 @@ and SFB. This leads to a GEMM $D = \alpha \text{SFA} * A \text{ SFB} * B + \beta
These can be represented in CuTe as:
- *SFA Layout*: $((\text{scale granularity M}, M / \text{scale granularity M}), (\text{scale granularity K}, K / \text{scale granularity K})) : ((0, int), (0, int))$
- *SFB Layout*: $((\text{scale granularity N}, M / \text{scale granularity M}), (\text{scale granularity K}, K / \text{scale granularity K})) : ((0, int), (0, int))$
- *SFB Layout*: $((\text{scale granularity N}, N / \text{scale granularity N}), (\text{scale granularity K}, K / \text{scale granularity K})) : ((0, int), (0, int))$
The 0 element stride ensures the same group of coordinates to map to the same element in the scale factors.