[TRTLLM-14702][feat] Integrate Kimi K3 KDA decode kernel - #17054

Merged

litaotju merged 10 commits into

NVIDIA:mainfrom

pengbowang-nv:dev-integrate-kimi-k3-kda-decode-kernel

Aug 5, 2026

Commits on Jul 29, 2026

[None][feat] Add KDA decode CUDA kernel and torch op

committed
c338b2d

Commits on Jul 30, 2026

add check and fix fake op schema
pengbowang-nv
committed
0d8f305
format code
pengbowang-nv
committed
dcfb5c9
apply kda kernel selection heuristic
pengbowang-nv
committed
d942283
fix output creation before shape validation
pengbowang-nv
committed
5ac7316
use int64_t for coord calculation
pengbowang-nv
committed
43a6a47
remove redundant dispatch code
pengbowang-nv
committed
dbbaf62

Commits on Aug 3, 2026

[TRTLLM-14702][fix] Guard cp.async paths in KDA decode kernel for pre-SM80 builds
brnguyen2
committed
5ba0c85

Commits on Aug 4, 2026

add check for kda decode input shape
pengbowang-nv
committed
9dc1796
enable test for other arch
pengbowang-nv
committed
5ab47f8