[TRTLLM-12807][feat] Wrap FP4 MLA as FMHA lib - #15375

Closed

yuxianq wants to merge 11 commits into

NVIDIA:mainfrom

yuxianq:feat/trtllm-12807-fp4-mla-fmha-lib

Commits on Jun 20, 2026

[None][feat] Add high-precision KV pool support for FP4 MLA

Tracin
authored and
yuxianq
committed
9743cf0
[None][feat] Support NVFP4 MLA KV cache on FlashInfer

Tracin
authored and
yuxianq
committed
7eec65c
[None][feat] Add CuTile and CuTe DSL FP4 MLA decode backends

Tracin
authored and
yuxianq
committed
c5c2232
[None][test] Add FP4 MLA decode benchmark

Tracin
authored and
yuxianq
committed
c1006af
[None][fix] Force CUDA utility paths for FP4 MLA debugging

Tracin
authored and
yuxianq
committed
5946c57
Support FP4 MLA linear MTP

Tracin
authored and
yuxianq
committed
2cb6f17
Apply FP4 MLA v2 integration fixes

Tracin
authored and
yuxianq
committed
f6d19cd
Remove debug codes.

Tracin
authored and
yuxianq
committed
ebd1347
Fix MTP + CUDA graph.

Tracin
authored and
yuxianq
committed
233176a
[TRTLLM-12807][feat] Add TRTLLM FP4 MLA FMHA lib
yuxianq
committed
a78bc11
[TRTLLM-12807][fix] Fix PR gate checks
yuxianq
committed
3549726