onnxruntime
[CUDA] Add an opt-in FP8 DeepGEMM MoE decode path for QMoE
#32122
Merged
Merged
[CUDA] Add an opt-in FP8 DeepGEMM MoE decode path for QMoE #32122
tianleiwu
merged
12 commits into
main
from
tlwu/20260816/qmoe_deepgemm
Add DSV4 DeepGEMM path (22.2% faster verify)
c4cd1dcd
refine deepgeem and update doc
1d5e5398
[CUDA] Skip padded rows in the DeepGEMM MoE pack and SwiGLU kernels
d93f0a5f
[CUDA] Use FP8 e4m3 weights for the DSV4 DeepGEMM MoE path
98bfc9c6
QMoE DSV4 DeepGEMM: make the expert count a template argument (32 and…
2f72c578
address feedbacks
5d78d939
github-actions
commented
on 2026-08-17
rename and update doc
13682197
lintrunner
30eebd9c
fix(cuda): preserve QMoE FP4 global scale
176e4cdd
Merge branch 'main' into tlwu/20260816/qmoe_deepgemm
72a49892
Merge remote-tracking branch 'origin/main' into tlwu/20260816/qmoe_de…
3b695d40
fix(cuda): address DeepGEMM review feedback
85eb5bb3
hariharans29
approved these changes
on 2026-09-04
hanbitmyths
approved these changes
on 2026-09-04
Reviewers
Assignees
No one assigned
Labels
model:dsv4
Milestone
No milestone
Login to write a write a comment.
Login via GitHub