leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 24 Packages Projects Releases Wiki Activity

v3.9 (#2185)

Browse Source 
* v3.8 update x

* fix blackwell gg

* doc change

* doc change

* doc change

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
Co-authored-by: Haicheng Wu <57973641+hwu36@users.noreply.github.com>
This commit is contained in:
Yujia Zhai
2025-03-21 01:52:23 -04:00
committed by GitHub
co-authored by yuzhai Haicheng Wu Haicheng Wu
parent 8c4d1dc47d
commit 62750a2b75
334 changed files with 91517 additions and 2656 deletions
Download Patch File Download Diff File Expand all files Collapse all files
media/docs/blackwell_functionality.md
+2 -2
View File
@@ -530,13 +530,13 @@ If the scale factor tensor exceeds M128xSF4, it indicates that there are multipl
<img src="../images/narrow_precison_multiple_block_sf_layout.png" alt="/narrow_precison_multiple_block_sf_layout.png"/>
</p>
The creation of scale factor tensors' layouts are tedious. CUTLASS provides `Sm100BlockScaledConfig` to create these layouts easily
The creation of scale factor tensors' layouts are tedious. CUTLASS provides `Sm1xxBlockScaledConfig` to create these layouts easily
(See [sm100_blockscaled_layout.hpp](cutlass/include/cutlass/detail/sm100_blockscaled_layout.hpp)).
The interface to create SFA and SFB tensor layouts is as follows:
```cpp
auto problem_shape = make_shape(M, N, K, L);
using SfConfig = Sm100BlockScaledConfig<SFVecSize>;
using SfConfig = Sm1xxBlockScaledConfig<SFVecSize>;
// SFA shape: ((32,4), ceil(M/128)), ((SFVecSize,4), ceil(K/4), L)
auto layout_sfa = SfConfig::tile_atom_to_shape_SFA(problem_shape);