feat(kda): integrate KDA attention with tokamax backend and CP support - #5128

Open

chiaotung97 wants to merge 14 commits into

AI-Hypercomputer:mainfrom

antgroup:feature_kda_integration

Commits on Sep 3, 2026

feat(kda): integrate KDA attention with tokamax backend and CP support
chiaotung97
committed
cac0254
fix: address PR review - terminology, guard removal, doc updates
chiaotung97
committed
cb947a9
fix: address remaining review findings
chiaotung97
committed
9e3cfce
fix: pyink formatting and doc consistency
chiaotung97
committed
77456fc
fix(kda): use mosaic implementation name, mark UTs tpu_only, add e2e smoke
chiaotung97
committed
233be2b
fix(kda): align layer and kernel API with final tokamax kimi_delta_attention
chiaotung97
committed
e6610e8
fix(kda): address review feedback
chiaotung97
committed
a368d8f
fix(kda): address second-round review feedback
chiaotung97
committed
d11e294
fix(kda): clear remaining lints and raise kernel-guard coverage
chiaotung97
committed
371f414
fix(kda): smoke delay must exceed the conv receptive field
chiaotung97
committed
85975d7
docs(kda): move KDA CP design doc into docs/reference and link it
chiaotung97
committed
402bc86

Commits on Sep 4, 2026