BREAKTHROUGH: Found the real bug!

Browse Source

Each epilogue thread owns only 32 of 128 columns per row (4 warps partition
the 128 columns). The scalar row_max is max of 32 elements (not global row max).
The scalar row_sum is sum of 32 P values (not all 128).

Evidence: kernel row_sum=5.5 for row 0, correct=29.22, ratio=5.3x
(128/32 = 4, but the P values arent uniformly distributed so not exact).

Fix: need warp reduce across the 4 threads that own the same rows columns.
CUTLASS FMHA uses utils.sm100 warp-reduce helpers for this exact reason.

...

This commit is contained in:

biondizzle

2026-05-21 19:10:47 +00:00

parent c09c660110

commit 8eb569e31c

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download Patch File Download Diff File Expand all files Collapse all files

Diff Content Not Available