[TRTLLM-12242][feat] Add Marlin NVFP4 backend for MoE and Linear on Hopper - #13476

Merged

shuyixiong merged 29 commits into

NVIDIA:mainfrom

xuantengh:nvfp4-hopper

Jun 25, 2026

Commits on Jun 24, 2026

[None][feat] Add Marlin NVFP4 backend for MoE and Linear on Hopper
xuantengh
committed
f998b21
[None][test] Add MMLU accuracy test for Nemotron-Super NVFP4 Marlin
xuantengh
committed
aec5574
enable multi-GPU support
xuantengh
committed
8354f09
update config
xuantengh
committed
5d295b4
add unified nvfp4 marlin check interface
xuantengh
committed
fa5fe63
refactor cuda / c++ kernel code
xuantengh
committed
0cb82b2
fix format
xuantengh
committed
19a22d6
Merge marlin test into existing NT3 super
xuantengh
committed
a123e31
fix compile error
xuantengh
committed
77480fc
fix linear test error
xuantengh
committed
f70fe5a
fix hopper only marlin kernel
xuantengh
committed
ca519ab
remove sa
xuantengh
committed
85cdd77
fix rebase error
xuantengh
committed
f2ab8f3
remove unnecessary configs
xuantengh
committed
dbbd283
add to qa list, and use new buffer protocol
xuantengh
committed
d070d7a
revise moe py
xuantengh
committed
b302bf9
reuse marlin into existing tests
xuantengh
committed
fdf3d06
align skill
xuantengh
committed
8d116a1
move quant method to dedicated file
xuantengh
committed
40fcf2a
add marlin to test_moe_module.py
xuantengh
committed
1a5e53a
reject marlin for configure moe
xuantengh
committed
64659e9
fix forcing triton
xuantengh
committed
88e6c01
resolve comments
xuantengh
committed
02140fd
add ultra
xuantengh
committed
4d43c13
add tests for nano/super/ultra
xuantengh
committed
556a63f
fix test name
xuantengh
committed
8de2a67
add abnormal shape to test
xuantengh
committed
3abeef0
remove unnecessary cofnigs
xuantengh
committed
e00b2e9
Merge branch 'main' into nvfp4-hopper
xuantengh
authored
77aaf06

Navigation Menu

Uh oh!

[TRTLLM-12242][feat] Add Marlin NVFP4 backend for MoE and Linear on Hopper - #13476

[TRTLLM-12242][feat] Add Marlin NVFP4 backend for MoE and Linear on Hopper#13476
shuyixiong merged 29 commits into
NVIDIA:mainfrom
xuantengh:nvfp4-hopper

Commits on Jun 24, 2026

[None][feat] Add Marlin NVFP4 backend for MoE and Linear on Hopper

[None][test] Add MMLU accuracy test for Nemotron-Super NVFP4 Marlin

enable multi-GPU support

update config

add unified nvfp4 marlin check interface

refactor cuda / c++ kernel code

fix format

Merge marlin test into existing NT3 super

fix compile error

fix linear test error

fix hopper only marlin kernel

remove sa

fix rebase error

remove unnecessary configs

add to qa list, and use new buffer protocol

revise moe py

reuse marlin into existing tests

align skill

move quant method to dedicated file

add marlin to test_moe_module.py

reject marlin for configure moe

fix forcing triton

resolve comments

add ultra

add tests for nano/super/ultra

fix test name

add abnormal shape to test

remove unnecessary cofnigs

Merge branch 'main' into nvfp4-hopper