[PyTorch][NVFP4][MOE] NVFP4 Grouped Hadamard Amax Kernel - #2351

Merged

timmoon10 merged 24 commits into

NVIDIA:mainfrom

zhongbozhu:zhongbo/mult_RHT_amax

Nov 25, 2025

Commits on Nov 22, 2025

minor fix of torch view dtype
zhongbozhu
committed
171dfbf
multi-tensor RHT amax, compiles
zhongbozhu
committed
6ec5dbb
setup multi_tensor_quantize_nvfp4_impl
zhongbozhu
committed
884bf36
wire things up and run without crash
zhongbozhu
committed
905b574
numerical test
zhongbozhu
committed
5515143
unit test passing
zhongbozhu
committed
84eba50
finish unit test of split quantize api
zhongbozhu
committed
a4f772e
bump up padding to 64 for nvfp4 grouped quantize
zhongbozhu
committed
6ac990c
fix stochastic rounding
zhongbozhu
committed
4e5c954
lint
zhongbozhu
committed
715cbce
change error message
zhongbozhu
committed
4e65dc5
clean up
zhongbozhu
committed
ea0fcda
enable multi-amax without RHT
zhongbozhu
committed
d3fadc3
fix col-only quantize mode
zhongbozhu
committed
0852c1d
improve benchmark script
zhongbozhu
committed
e5e6209
add NCU example script
zhongbozhu
committed
dd887ac
add larger test case
zhongbozhu
committed
6a3893a
add contiguous_data_and_scale check to bulk allocator
zhongbozhu
committed
07dc9b3
unified naming and differentiate between group_ and multi_
zhongbozhu
committed
e05a11b
move regular amax into multi_tensor.h
zhongbozhu
committed
a2a7365

Commits on Nov 24, 2025

Disentangle logic for split-quantize and general multi-tensor quantize
timmoon10
committed
16017c6

Commits on Nov 25, 2025

Use size_t for split sections
timmoon10
committed
679f9a4
Suggestions from @greptile-apps
timmoon10
committed
419a2b9
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
44f43d5