Files
nvfp4-megamoe-kernel/tests
biondizzle e85d50dc3b fix: compute row_max from RAW S values, not scaled
row_max should be the max of the raw QK scores, not pre-scaled.
The scale_log2 is applied during exp2 and rescaling, not stored in row_max.
This fixes the double-scaling bug that broke multi-tile O rescaling.
2026-05-22 10:21:50 +00:00
..
2026-05-22 08:57:38 +00:00