[None][fix] Align the budget split, spec layers and pool_ratio with derived per-layer KV windows - #18697

Merged

brnguyen2 merged 2 commits into

NVIDIA:mainfrom

moraxu:fix/derive-per-layer-attention-windows-from-layer-types

Sep 16, 2026

Commits on Sep 14, 2026

[None][fix] Align the budget split, spec layers and pool_ratio with derived per-layer KV windows

moraxu
authored and
brnguyen2
committed
10c8d85

Commits on Sep 15, 2026

[None][test] Update budget-split expectations for the corrected V2 quota math
brnguyen2
committed
aa614b3