biondizzle d7aa4da686 BREAKTHROUGH: unnormalized P@V cosine 0.999998 for n=128! ...

The softmax math (exp2, P store, PV) is correct for single-tile.
The bug is ONLY in C6/C9 normalization: applying inv_row_sum
using PV partition instead of QK partition.

n=128 (single tile): cosine 0.999998 PASS
n=256/384 (multi-tile): C6 O-rescale using wrong partition = FAIL

Fix: normalize O using QK row coordinates, not PV row coordinates.
Can use TMEM vector to bridge QK partition to PV partition.

2026-05-21 18:55:00 +00:00

..

archive

Restructure: cutedsl/ -> dsv4/ with proper layering

2026-05-21 17:30:44 +00:00

e2e

Restructure: cutedsl/ -> dsv4/ with proper layering

2026-05-21 17:30:44 +00:00

integration

Restructure: cutedsl/ -> dsv4/ with proper layering

2026-05-21 17:30:44 +00:00

unit

BREAKTHROUGH: unnormalized P@V cosine 0.999998 for n=128!

2026-05-21 18:55:00 +00:00

requirements.txt

test: add standalone layer 0 comparison test (no vLLM, no Docker)

2026-05-16 02:13:18 +00:00

run_test.sh

fix: use setup.py install for CUTLASS extension build

2026-05-16 02:21:17 +00:00