biondizzle 9ea5551241 FMHA Stage-C: real softmax + O normalization in 6-warp layout ...

- Replace identity softmax with online softmax (row_max, exp2 scaling, P store)
- Add row_sum accumulation from P values
- After softmax loop, normalize O in TMEM by 1/row_sum using TMEM load/modify/store
- Then epilogue writes normalized O from TMEM to GMEM
- Reference test uses softmax(Q@K^T/sqrt(d))@V

2026-05-22 09:22:56 +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

FMHA Stage-C: real softmax + O normalization in 6-warp layout

2026-05-22 09:22:56 +00:00

native_stage_c_patch.py

more stuff

2026-05-22 08:57:38 +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