biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

D1.5 debug: try corr_tile_size=32 for O rescale round-trip

Browse Source
This commit is contained in:
biondizzle
2026-05-26 20:43:29 +00:00
parent 20ed6d5114
commit e35b30dae6
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
dsv4/kernels/attention/fmha.py
View File

@@ -418,7 +418,7 @@ class FmhaKernel:
# sub-tiles, and building BOTH copies from the SAME tensor, ensures the
# column mappings agree on round-trip.
# ============================================================
corr_tile_size = 16 # Must be power of 2, divides head_dim
corr_tile_size = 32 # Must be power of 2, divides head_dim. Try 32 instead of 16.
tOtO_i_layout = cute.composition(
tOtO0.layout, cute.make_layout((128, corr_tile_size))
)