[None][feat] Two-level GVR decode top-K dispatch; remove the CUDA heuristic and temporal-only prior state - #18446

Merged

juney-nvidia merged 13 commits into

NVIDIA:mainfrom

longcheng-nv:feat/gvr-unified-router

Sep 8, 2026

Commits on Sep 3, 2026

[None][feat] Two-level GVR decode top-K dispatch from the sparse-attention config

longcheng-nv
and
claude
committed
7c92a3b
[None][refactor] Remove the CUDA GVR heuristic top-K decode

longcheng-nv
and
claude
committed
d690470
[None][refactor] Skip GVR prior state for the self-sampling engine

longcheng-nv
and
claude
committed
aaba6b0
[None][feat] Emission block-skip as a third GVR decode top-K dispatch param

longcheng-nv
and
claude
committed
9aad420
[None][test] Regenerate the LLM args telemetry golden manifest for the GVR dispatch fields

longcheng-nv
and
claude
committed
59d051a

Commits on Sep 4, 2026

Commits on Sep 5, 2026

Commits on Sep 6, 2026

[None][infra] Restore blossom-ci allowlist to match main
longcheng-nv
committed
b3a4fb2