[NVFP4][Dense/MoE] Integrate Cutlass NVFP4 Row-Cast-Col-RHT-Transpose-Cast Fusion Kernel - #2555

Merged

ptrendx merged 28 commits into

NVIDIA:mainfrom

zhongbozhu:zhongbo/dense_row_col_rht_fp4_quant

Mar 16, 2026

Commits on Mar 12, 2026

first draft
zhongbozhu
committed
ee1e4db
pass numerical unit test
zhongbozhu
committed
0efd73d
format
zhongbozhu
committed
bf58336
add benchmark script
zhongbozhu
committed
3513581
lint and format
zhongbozhu
committed
5f0c595
compile guard
zhongbozhu
committed
98c30cf
warning fix
zhongbozhu
committed
d2e360d
resolve greptile comment
zhongbozhu
committed
f667cb1
minor style fixes
zhongbozhu
committed
b6cd482
fix namespace
zhongbozhu
committed
7801586
resolve some comments
zhongbozhu
committed
1df7de0
fix comment
zhongbozhu
committed
8e2023a
attempt to fix compile CI with guard
zhongbozhu
committed
0bc8a35
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
zhongbozhu
committed
1a29e4c
better naming for tests
zhongbozhu
committed
32948ab
fix deprecate messsage
zhongbozhu
committed
f324eac
more compile guard
zhongbozhu
committed
953e09d
new API name
zhongbozhu
committed
a004811
fix format all in one
zhongbozhu
committed
b1a8d34
try to fix compile CI again
zhongbozhu
committed
81ac960
AI code review comments
zhongbozhu
committed
113f940
to pass oldest compile CI with cuda 12.1
zhongbozhu
committed
0de8031
add more guards to nvfp4
zhongbozhu
committed
1498ca8
make multiply inverse default numerics
zhongbozhu
committed
950303b
update numerics of nvfp4 partial cast as well
zhongbozhu
committed
4dce48b
resolve comments
zhongbozhu
committed
b13e518
add NVTE_BUILD_NUM_PHILOX_ROUNDS after rebase
zhongbozhu
committed
f6ddc6a
simplify compile guard messsages
zhongbozhu
committed
6dabcee