NVIDIA/cutlass
1
0
Fork 0
mirror of https://github.com/NVIDIA/cutlass.git synced 2026-05-12 01:10:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

Small Tile N BlockScaled GEMM + Grouped GEMM (#3176)

Browse Source 
Co-authored-by: dePaul Miller <23461061+depaulmillz@users.noreply.github.com>
This commit is contained in:
dePaul Miller
2026-04-21 09:32:40 -07:00
committed by GitHub
parent aeba0d3723
commit b46b16d003
13 changed files with 468 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
python/cutlass_library/generator.py
View File

@@ -11192,6 +11192,8 @@ def GenerateSM120_TensorOp_mixed_8bits_UMMA_gemm_with_block_scaled(manifest, cud
]
tile_sizes = [
[128, 32, 128],
[128, 64, 128],
[128, 128, 128]
]
@@ -11324,12 +11326,20 @@ def GenerateSM120_TensorOp_fp4_UMMA_gemm_with_block_scaled(manifest, cuda_versio
]
tile_sizes_cooperative = [
[128, 32, 128],
[128, 32, 256],
[128, 64, 128],
[128, 64, 256],
[128, 128, 128],
[128, 128, 256],
[256, 128, 128]
]
tile_sizes_pingpong = [
[128, 32, 128],
[128, 32, 256],
[128, 64, 128],
[128, 64, 256],
[128, 128, 128],
[128, 128, 256]
]